The Never Ending Lore of Harness | Vivek Trivedy (Product Lead, Langchain)

Hey everyone, welcome back to ground

zero. This is episode 13. Yeah, we are

running fast. Today we have ve from

langchain. So we leads their work on

open source agents and harnesses the hot

term right now. He's the person behind

DP agents the coding agent that went

from top 30 to top five on terminal

bench 2.0 by only changing the harness.

He's been writing some really good stuff

with lot of signal and alpha on what

harnesses actually are. Why agents

should be more opinated the idea of

harness as a service and um how planning

agents are really just dynamic workflow

generators.

Before Langen, he ran his own startup on

visual understanding agents and before

that uh was a scientist at AWS while

doing his PhD in CS at Temple. Uh, we'll

cover a lot into this. Uh, there's a lot

to get into. We welcome.

>> Thank you for having me. I'm super

hyped. I'm super I've been following you

on Twitter a bunch. So, yeah, I'm glad

we're making this happen.

>> How are you doing? And would love to

know your uh initial VIP check on Opus

4.7.

>> First of all, doing great. Whenever

there's a new model release, you know,

it's always like a good week for all of

us. It's maybe like an even more fun

week for like anyone who does like evals

on all the models. Um, so yeah, dropped

yesterday. We started like evaling it.

We have our like set across our

products. We have like open source evals

that we use and like also like for some

of like Lang Smith's products that we

use. It's a good model. It's a good

model. I don't think it was like a crazy

step change for tons of stuff that we're

doing. But TBD I think like the fun part

about stuff we'll like jump into which

is strong belief that every model needs

its own custom things that you add to

it. I know like anthropic release is a

nice skill uh that you can like easily

convert prompts and stuff but we're in

the middle of that process for like the

agents that we're going to use it for.

So it's a good model not a crazy step

change but we'll we'll fit it. We'll

we'll make it good.

>> I mean it is interesting in a way that I

have been seeing a lot of mixed opinions

right now. People have pretty much mixed

opinions on 4.7. Basically what they

have doing it with um the kota users as

well. I mean in just three four prompts

you are running out of I mean there's a

lot of good story I mean interesting

story behind but but yeah I mean the

kind of piece about these models being

coming up be open air or anthropic

anthropic specifically how they have

been doing good at public perception and

effective marketing as I say I mean

working well working working I mean it's

been rewarding for them

>> I mean they're great they're great they

they put out like great models obviously

they put out great products around the

models I think there's definitely some

stuff where

people are playing a lot more with the

models and like they're basically like

picking use cases they use models for.

So it's like everyone uses cloud code

like everyone uses codecs and that sort

of stuff. But like when you build like

your agents on top of those models, it's

like I need to actually care about the

prompts. I need to care about the

context engineering. I need to like care

about the tool design. And I think like

that's where it's really cool to like us

putting out content like other like

really cool people putting out content

which is like like how do I make a model

good at like my task basically because

at the end like my customers that's all

they care about that's all I care about

and I think like that's like a bunch of

the harnessge journey basically whether

you call context whether you call like

agent edge it's basically like fit some

sort of system around this model to make

it like sit at my task and that's like

what we're all trying to do and like

anthropic is trying to help us with

that. Open models are trying to help us

with that as well.

>> Totally makes sense. Um let's dive in um

about your journey. So you went for a

PhD in CS at Temple and I mean worth to

mention you did your bachelor's,

masters, PhD everything at Temple and

this has been a talk of the town as well

in past years on Twitter. People were

talking about it. People have again I

mean some opinions about Temple being a

university, good university or not. So

my question is to being a scientist I

mean doing a PhD PhD then to being a

scientist at AWS to running your own

startup on agents or visual

understanding to leading open source

agents at Langen. How has your journey

been like?

>> Happy to dive in. Um cool cool I'm so

I'm from around this area. So I'm from

like east coast uh Jersey like

Philadelphia area. I went to school at

Temple. So I did my undergrad there did

my masters there like my PhD there. So

like super early I was like I'm just

going to be a doctor like most kids

pressured by their parents like I'm

going to be a great doctor like quickly

realized like I don't really want to do

that most of my undergrad. So I do my

underground in math and math is like

really cool. I think there's a lot of

concepts in math that like translate

really well to CS and like physics and

things like that sort of like systems

thinking.

>> Math is also like at least for me maybe

I'm just not amazing at it. It's

incredibly hard. So like doing something

really hard does prepare you for other

things.

Yeah, dude. Undergrad was like really

fun. I enjoyed math. I got into like

some CS stuff. I think like late 2010s

was when there was a lot of cool stuff

in different parts of ML. So like I got

into computer vision stuff, like

undergrad research. And like I love

vision. So like I think vision is still

one of the coolest things out there.

There's like way less research done on

vision even today relative to text. Like

>> OCR is pretty important, right? OCR is

like now okay just just send just send

the PDF to Claude basically and like

obviously a bunch of systems engineering

around that but yeah man like I I loved

vision I still love vision vision was

really cool so like I did undergrad in

that did like research around that and

then I just went straight into like

masters in PhD like right after I

graduated like early 2020s and then yeah

my PhD was basically all around like

vision focused representation learning

so yeah I can talk a little bit about

that. So the first like topics that I

was working on was like graph neural

networks which are like I don't know how

hot those are anymore but I do see like

some really cool people still doing

research around those. Um basically like

graph representation learning but it's

like graph representation learning for

like vision basically. So it's like if I

like decompose an image into like

particular objects and like I make a

graph of that and then I do like

representation learning do we get like a

better end vector for like retrieval

like classification and then like we did

this at also like the data set level as

well. So like what if I have like kind

of like few shot examples. It's called

like transductive learning like use

other information in the data set to

help you classify the next thing. Dude,

that was really cool. Like I think

graphs I'm like bearish on graphs

overall actually. So maybe hot take but

like that was a really cool part of

research and like that was my first like

dabbling into like computer vision stuff

like undergrad then my first like PhD

topic which like it shifted a little bit

after like the chat PT moment like tons

of research became around okay like

let's do VLMs for everything and let's

do like representation learning on the

VLMs like what are VLMs like actually

seeing when they're doing their like

attention mechanism over images. So

yeah, dude, it was great. It was great.

I like really enjoyed my time in PhD. I

think it's like you get some sort of

unbounded time with your adviser to just

pick an interesting problem and just

like rabbit hole in it. So I did like

retrieval stuff like representation

learning stuff. Yeah, dude. It was

great. I enjoyed it.

>> Awesome. Um, so I had a chat with

Tensorcut the other day. He started

Paradigma. He dropped out of PhD. So my

question to you is what do you really

think about the scenario right now the

linkage between academia and the

industry and right now if you have been

like if someone is going for PhD or

something like that. So what do you

really think about is is is that is this

worth it or how far we have come is

still necessary to go for a PhD to I

mean it is again very opinionated um

question but still I mean I want to

really understand your

>> yeah absolutely so like it's a great

question like people ask me this

question like locally like my friends or

like younger brothers and stuff.

>> Yeah. So like maybe my PhD was like

slightly different because I was doing

research at Temple but I was also doing

research and like working on like prod

projects when I was at AWS and those are

happening at the same time and I like

strongly believe that that is like a

fantastic mix for anyone who wants to do

like research but then sort of

understand maybe like how their research

is going to be applied in like some

settings. And I think today like the

point basically of a PhD to me is like

you pick a topic that you're like really

deeply interested in and you like poke

around the edges of that topic to try to

figure out like how we can make like

this thing better. And like that doesn't

like really require a degree to do that.

There's tons of like sick researchers on

X who just like post like random blogs

and like they don't have a PhD. they

probably don't maybe don't have CS

background but there's like you just

pick a topic you like rabbit hole it

you just like push the boundary of

what's possible and you do that like in

a verifiable way so you like write code

do experiments you try to share like

open research and if you're able to find

a company that allows you to do that

like lang's fantastic at that like I

think they really cultivate like hey

like we're going to like pick this topic

we're just going to like figure out how

it works and we're going to like publish

content about it basically

>> I would say that's great I think it it

kind of depends like if you find a great

company, a good great founder that you

vibe with that lets you do both.

Industry is like amazing and like

especially AI research like it's super

helpful across a lot of companies. You

can probably make a lot of money and

like do interesting research at the same

time. So yeah, kind of like a

non-answer, but if you do find that

scenario amazing if you just want to

like grind on like some sort of topic

and PhD for like a bunch of years, also

great. I actually don't think you can go

wrong like just by being curious and

just exploring it.

>> Yep. I can see you have uh you you were

like working on your startup about

visual understanding agents. So I want

to understand your learnings there and

how do you see the vision space right

now like how can you correlate between

uh the time when you started and the

time we have come so far with the

current frontier state-of-the-art

research and products building. Yeah,

dude. Um, yeah. So, like I started that

startup after I graduated like my PhD.

So, that was sort of like mid last year

with a friend. And basically like the

main thing that we were working on like

starts was called Agentify. And like the

main idea was basically that basically

vision compared to text like really lags

behind in frontier models for like

things like visual reasoning but also

things like perception just generally.

So there's like tons of things where

you'll like show an image or like an

object like o two overlapping boxes to

the model, right? And it's like it

doesn't like fully understand that those

two things like overlapping and like

part of this is just a perception

problem in the visual encoder where it's

like some of these like fine grain

details, it's just not able to

understand them with like the native

training that it has. But that I think

is like a fantastic opportunity because

it's like how much of that gets absorbed

into the vision encoder backbone versus

like how much do we augment models with

like tool calling behavior that they're

exceptional at and actually use that as

the mechanism to like take vision

capabilities and like put them into the

models. Like that's basically the whole

like idea that we were working on. So

like research and like product around

that which is like what if I just took

all of the classic vision models that we

already have and like a lot of this was

honestly inspired by Meta's work on SAM.

So I think like SAM and that whole

series is like incredible like SAM 123.

It also supports like video segmentation

which is like insane and you can also

like fine-tune it. You can do like meds

SAM and things like that. So it's

basically like BET was okay models are

amazing. They're getting very smart, but

like their vision capabilities are

lagging behind. But we can augment them

with tools and like you can basically

like do the right tool selection in the

moment to like get that capability. Like

segmentation is something that it was in

Gemini Flash across the Gemini series,

but like compare that to like SAM,

right? Like SAM was like way better. If

you just use like Sam as a tool compared

to like the native segmentation Gemini,

you would be just like way happier. and

like all you had to really do was like

point to the right spot which is like

way easier than doing like semantic

segmentation. So that was the idea. I

still think that that is true in vision

today. Like even with like Opus 4.7's

new benches, it's still not as good at

visual perception as like we need it to

be. So I still think tool use is like

really really exciting for yeah just for

like agentic systems like visual

basically making a bunch of like vision

specific tools for your task and like

augmenting uh yeah augmenting your agent

with that.

>> I think there is a lot of scope to do

alongside UI bench as well. I mean again

uh it's more about one's taste but uh

but there are lots of ifs and buts lot

of nuances where you really need to take

care of like even if you're cloning a

website I mean there's lot of sc uh

scope to play around something so my

next question is about your work at

Loheed Martin. So you you you interned

there. I think that was your first um

job and honestly a lot of what people

see about world is kind of sophisticated

reals on social media about American

weaponry. So what was the reality like

from the inside? What what what you were

working on? How does it feel like to

work at some defense um kind of defense

company and what experience lead?

>> That is like such a throwback. So that

was like my first internship at like

tech ever. So, I was like a bio intern

in undergrad and I was like looking for

internships and I gave my resume and I

got an internship at like Loy Martin

which is amazing because like I don't

know how good my bio resume was for

getting like any internships. Yeah, man.

I wish I say like tons of stuff I did on

>> What do you mean by bio resume? It was

like like you were working on some bio

>> Yeah. So like I went to undergrad as

like a biochem major because like I

wanted to be like a doctor.

>> Amazing.

>> Yeah. So like then like after freshman

year I applied to like internships cuz I

I switched I wanted to do tech after

that or like at least explore it with

like a bio resume and they were like

dude like what like what are what are we

doing here? And then like I think I

basically just like talked like to the

hiring manager and just said like hey

I'm like really down to like learn this

thing like which is like data science

like that time there bunch of these like

data science courses and things coming

out so it was still like early and I was

like hey like I took these like Python

classes and like I'm super down to learn

this. And basically it was like yeah I

mean it sounds great.

I ended up working on the data science

team there and it was basically like my

first introduction into like kind of

like data analysis sort of stuff. So

like understanding like it was much like

stats basically. So like I wouldn't say

it was like ML but it was like this is

like intro to like making plots like

slice this data this way. So it was a

bunch of just like empathy for like very

very messy data as like my first

internship which is actually like very

valuable today just like insane amounts

of data which is like does not look very

clean and yeah man I wish I say more it

was basically like a great learning

experience because I was kind of

learning how to code and like doing like

data science stuff and then it was also

like a decent confidence boost because

I'm like okay maybe I can do like tech

stuff and yeah I interned there and it

was like fun and then yeah I didn't

really go back after that but I started

getting into more like research stuff at

school.

>> Awesome. Um, also recently I was just

kind of exploring the timeline. I see

Mike Mill who is a pretty famous, you

know, internet celebrity was looking for

an AI guy and you came up through

Temple. Apparently Mike was surprised

how many Temple people are in AI and so

did you end up connecting with him? Did

you share anything about Langchen and

stuff?

>> So Meek Mill is like he's like a rapper

from from Philadelphia and like I guess

he lives around Temple like that's where

he was from and I think everyone was

like when they saw that tweet they were

like Meek Mills get into AI so okay let

me just like reply basically because I

think like honestly like randomly

posting on Twitter X is like awesome.

You can meet so many cool people like

that and I we'll talk about this but I

met like Harrison the founder of like

CEO and the CEO of like W

And yeah, he did not reply to me. I hope

his like startup is doing sick, whatever

he's whatever he's doing. But like I'll

like repeat it if he does need someone

for help with like AI. I'm actually like

seven blocks down. So I could totally

like just pull up and help him. So no, I

think that's a good lesson though is

just like randomly posting maybe like

I'll just keep doing that and then maybe

something will happen.

>> Yep. Awesome.

So I mean the next question to you is so

when did you join Langchain and uh what

actually pulled you there specifically?

So and since you joined what actually

has

>> So this is like this is so much fun. Um

I was working on my startup like after I

finished my PhD that didn't work out

like we basically stopped around the

fall. At the same time, I was basically

like doing my first foray into just like

posting like random stuff on Twitter

just like my thoughts like basically

just like open source stuff like hacking

on random stuff and

from a bunch of the stuff I was posting

around like so like last year I also

like sort of believe that like we have

amazing models but like because we did a

bunch of stuff in this like visual

understanding space with like agents and

stuff. I was like very very confident

that models need like some stuff around

them to like help them do these tasks

because like they just suck at them out

of the box and like we basically saw

this every day. So that's basically when

a lot of maybe the ideas that were

brewing around harnessge like started to

maybe get more like crystallized and I

just started like posting about that

online. It's like, hey, like this is

maybe like what harnesses look like.

Like harnesses are like supposed to like

wrap models and like if we're trying to

do like vertical tasks. It like really

helps to have some sort of like

opinionated like prompts, context

engineering, like tool call structure

like all this sort of stuff. And I think

I just like started DMing Harrison like

the CEO from that which is like super

sick. He is also always thinking about

like the frontier of like AI systems

which is awesome. And then we started

chatting maybe like late last year just

like yeah like what would it look like

to build open-source infrastructure

around like agent engineering and like

maybe the best way to facilitate that is

by helping people build good harnesses

like whatever good means like let's

discover like what good means and make

open source software about that. So, it

was basically like, okay, that sounds

sick. And then I was like, I don't

exactly know what I'm going to do. Like,

maybe I'll continue like working on the

startup or like, but I would love to

join something that like really aligns.

So, then I started working with like

their open source team late like last

year on what ended up becoming like what

was deep agents, but ended up becoming

like a lot bigger. Um, so yeah, we were

working on like the very very early

versions of like deep agents last year,

which is like one of our libraries at

Langchain that we that we have. It's

like our library to help people build

harnesses. Um, or at least it's one of

the ways that people can build harnesses

using using Wangchain. And yeah, I loved

it. I love the team. Uh, amazing people

doing open source. And then I decided to

join like full-time in in December.

>> Amazing. Um, and and I mean, the

adoption is just crazy, dude. I mean, so

I want to understand about the growth

here. So, so again I mean right now

Twitter is full of people flaming

millions in ARR every month and but like

a feels like one of the most you know

stressed metrics out there. So my

question is how has lang approached

growth in real terms be it opensource be

it community adoption be it enterprise

or

>> yeah dude it's a great question. So I I

think about this a bunch because like I

think the best way to maybe think about

it is like basically like work backwards

from you want to like help people build

stuff using like the tools you're you're

putting out there, right? And like the

goal is basically just like help people

build like really cool things and like

make that process of building as easy as

possible. I think in like open source

that comes through like very clearly

because in open source I think you get a

lot of like empathy for the end user

because they're like directly using your

product like all the code is like fully

visible like go inspect it also like put

your opinions in like our GitHub issues

and tell us like what's good what's bad

like what should we fix like what should

we add also like it's totally cool to

like disagree in open source because

like the maintainers sort of have

limited bandwidth to address like all of

the things, but we want to make sure

that the most impactful things that are

going to help like the most users build

like the coolest stuff like we like

prioritize those. So, I think there's a

there's a big part of growth which is

why I like really like X um and like

these direct feedback channels or like

Slack for example or just like messaging

builders and customers because you

basically get to see exactly what

they're doing. you build like a lot of

empathy for shoot like this thing that

we built like it's a little broken in

this way or like it doesn't exactly like

fit the use case and then you hear a

bunch of those stories and you sort of

like work backwards to say okay like we

need to improve like this part of our

library or like we need to like make it

possible for others to improve our

library as well. That's like an amazing

part of open source that we get tons of

like amazing feedback, tons of like user

contributions which is great because you

sort of like grow with your community

and I think like that's a really big

part of open source and related to that

which I really really like about

Langchain like one of the reasons why I

joined and like I really enjoy working

here is there's a lot of like learnings

that we get from all the research that I

do in like open source and like putting

stuff out there and getting feedback

that slowly like make their way into our

products as well because it's like for

example a lot of stuff in like Lang

Smith for example which is like okay

like how do you build good evals like

how do you how do you actually enable

agents and users to build like really

good evals like how do you like

understand what's happening in traces

like mind signals from

>> like a lot of that we put out just in

the open like I did a bunch of blogs on

that stuff there's other people who are

like hacking on that stuff as well and a

lot of the stuff in open source you sort

of see how the community interacts with

it. You also just see the raw numbers

and you put it out there and it's like

hey like I would love this or like I'm

using this and it's like oh we should

make that as easy as possible. Put it

into a product and like if people love

the product then like the rest of it

sort of takes care of itself. It's like

yes you will make money you know your

customers will be really happy and then

like just continue the loop like just

keep making it better basically. So I

think like yeah dude customer feedback

is amazing like community feedback is

amazing. So it's like a really really

big part of I think lang chain a really

big part of like a a lot of the open

source stuff that we do

>> I can imagine of course and more

specifically here so you are leading the

open source egen and harnesses work

right now so what does a typical um week

looks like for you it's more about

research engineering or product

>> yeah dude whatever

>> I think the fun part is like it's it is

actually like a mix of a ton of stuff

and I like really really like that so

it's like the goal is bas basically pick

the most important thing to work on at

this time and then like we'll like we'll

chat about it maybe over the weekend or

like the week before like Harrison

jumped in with with us like we'll DM and

let's just like sprint towards that and

build it basically and like maybe what

that looks like lately

like lately like a ton of my work has

been on like eval continual learning

essentially like methods for using like

evals and continual learning to make

like agents and like their harness

better. So that's like basically like

the research direction and I would say

maybe like 50% of the week goes into

okay let's like pick a research

hypothesis let's like figure out what

the experiment design around that might

be. Like for example, last week we were

doing a bunch on can you like just in

time generate evals uh like for any

given task like what does that look

like? Like are you overfitting to them

and like what is your like fitting

algorithm? There's like tons of stuff

that we put out. There's like a lot of

good content on like harness hill

climbing basically. But yeah,

essentially it's like research. Let's

pick that task. Um kind of like a PhD.

We're going to make a hypothesis. We're

going to like run the experiments on it.

We're going to get like get metrics and

we're going to post them on Slack and

we're going to like review them and like

argue our takes about them essentially.

Yeah. Then the other maybe bunch of

percentage like 50% is like talking to

customers like talking to people like on

Twitter getting a bunch of feedback from

them on like the open source stuff like

how can we improve our libraries whether

that's like lang chain lang graph like

deep agents anything in like lang and

then a bunch of that is talking with

like product teams as well. So there's

like tons of great teams at Lang Chain

that do a bunch of good work on like all

the products that we have. So there's

tons of learnings that I think come from

open source that we can like port back

into the products that we're going to

build and yeah just keeping that

feedback loop is good. So I would say

like it's a mix bunch of like research

and then engineering stuff and then a

bunch of like I don't know like what the

term today is but like devril like

devril devx which is just like if

someone asks a question on Twitter like

we should respond to them and we should

like put our ideas out there and we

should like be willing to engage with

other people's ideas and yeah just hear

what people are saying. So it's like a

mix yeah it's a mix of those things.

what percentage of your article source

like article is coming from this

research source I can imagine a certain

percentage but because dude I mean I

mean let's just come to harnesses like

what this what is all about the load

behind harnesses right you know

>> so you mentioned that the definition of

agent is basically model plus harness

right

>> so I mean this is something like I mean

it is being in like people know this

from quite some time like this is this

is a fact but I think this is the

cleanest framing anyone any anyone have

seen at least on Twitter. So if you're

not the model, you are a harness, right?

And and a harness is every piece of

code, configuration or execution logic

that isn't the model itself.

>> So can you walk me through how you

arrive at the definition?

>> Yeah. Yeah. Yeah, dude. I think like it

is it is definitely like a cleanish sort

of specification of like what is this

thing that we're talking about and I

think like maybe the definition doesn't

really matter like as much like what the

exact equation is but like there is one

thing that's helpful which is like when

you're communicating with someone about

like how we're going to make this agent

better we need like some shared language

so we can talk about like what is the

thing that we're going to optimize

basically right so it's like

like working backward from model

capabilities because like that's sort of

the thing that we need to wrap

intelligence like wrap systems around to

like amplify the intelligence of the

model. So it's like I basically view it

as there's some sort of computation

happening inside the LLM and like where

that's happening is over this like

context window boundary. So like all the

compute happens when I basically like

take context from like my system and I

push it over the boundary and I put it

into the context window like for the

model to do computation on and then

produce tokens basically. And like some

of those tokens correspond to like tool

calls and then I go and execute those

tool calls and like I return the context

back. And like the reason why I like

that is because like models by

themselves they're basically just like

>> token input machines and like token

generators basically. But like we need

to put a system around the model so it

can do useful things. And I really like

maybe like working backwards from what

should the agent do and like maybe even

like what does my customer want the

agent to do and then like figure out if

I just like give it like a really really

simple model like maybe like really

really simple harness. Can the agent can

the model and like the agent can the

agent basically just do that? And like

if the agent can just do that with like

a really simple harness, then that's

like amazing because then we can just

like give that to the user essentially.

Where things maybe get like more

interesting is like where like a really

simple harness just like can't do that

today. And that might just be because

like it doesn't have the right tools or

maybe like the model isn't intelligent

enough to like orchestrate those tools

in order to do that. Or maybe it's like

some of our context engineering opinions

in the harness aren't good enough and

it's like hey like you're you're putting

a bunch of like really big tool call

outputs like into the context window and

it's like confusing the model. We should

find out ways to not do that. But these

are all basically like harness level

configurations that we're doing and

they're external to the model. Like the

model is basically just like a

computation unit and it computes things

over its context window and like we need

to decide what goes into that context

window so it can do like useful work for

us.

>> If I have to ask you some like three uh

three bullet points what really makes a

good hardness according to you what are

they?

>> Yeah. So there's a bunch, but if if I

had to pick like three right now, I

would say

basically prompting and like very very

clear instructions

for better or worse. Like there was this

whole thing like prompting is dead. Like

prompting is like totally not dead. It

is like so useful, so helpful. And like

I I don't just mean like prompting in

terms of just a system prompt. Like

prompting also applies to like the tool

descriptions as well that get like

autoloaded into context. It also applies

to how well your like skills front

matter explains like how to use these

skills or like how to use like other

skills. It it also applies to like if

you have sub agents, does like the sub

agent front matter specify like when

this should be used or like how to use

it basically. So it's just like

basically prompting that encodes really

really good instructions from the user

or on behalf of the user for like how to

use this agent to do useful work. That's

like super important. I think like

prompting is honestly more important

today than it ever was before because

our like the systems we have are way

more intelligent. So we're able to guide

them towards doing useful work more

easily with good prompts. That's one. I

think the other one that we're spending

a bunch of time on right now is

basically verification. So we did like

some blogs around this on like making

coding agents better. But there's sort

of like maybe two things in

verification. like first is prompting,

second is like verification. So there's

like a built-in verification that you

might inject like into into the harness

itself. So like that can be like a hook

basically. So like before the model

tries to go and exit like force it to

like recheck the work or like make sure

>> really

>> verification is basically like if if I

give so for example if we just use like

all the terminal bench tasks, right? So

like terminal bench task comes with like

an environment. It comes with like a

task and then it comes with like a

verifier that will run after the agent

thinks it's done, right? But like

obviously we can't use that verifier

information. So like what the agent

needs to do is like it needs to like

self-verify its work before that

verifier runs to like be like very very

sure that the code that it developed

solves the task that we're that we're

like trying to solve. Maybe there's two

parts of that. One part is we need to

like teach agents what the useful

primitives are for verifying their work.

I think like one immediate one if like

anyone uses like the claude model or

like even like GPT 5.4 is like agents

are very susceptible towards like

picking the easy way out in verification

which is like they test like trivial

cases or like not not like very

difficult cases. Obviously, that fails

in the verifier because it's just like,

hey, like I checked like these three

cases are really easy, so like I'm good

essentially and like that's bad. Like we

should teach agents to be much more

thorough when they're like generating

verification for themselves. That's like

one part of it. The other part of it is

like like this is all code. So like we

have in our repos tons of like unit

tests and like tons of like evals that

we already use. Like that is great

context that we should give to the

agent. so that it can like run that eval

suite and that might be run with a hook

for example like I don't want like maybe

the agent won't run it by itself but

like when it tries to exit that should

just maybe run my eval suite or a subset

of it and it should inject the context

or like the results back to the agent so

the agent can see like what failed like

what what passed basically because like

we need some sort of signal to give back

to the agent so we can like fix the

thing that it generated so it's like

self-verify or like use external signals

from like existing evals so you can like

fix the things that are going wrong. And

I think that's like a really really big

part of it. And like maybe the last part

that we're focusing a ton on is

high level. It's kind of like

orchestration basically but for doing

things that are more long horizon

basically like it's problem

decomposition and like making sure that

like when we use like sub agents to do

problem decomposition like two things

are true. So one is we're picking the

right model like agent for the job

because like every model is like good at

different things and also that um this

is a lot of context engineering. We're

basically like bounding the sub problem

that the agent needs to do in like a

decent enough window that it can like

manage it. Basically what I mean by that

is um I wanted to like do things in like

a 50k to like a 150k token range roughly

or like 200k. sort of it depends on the

model but like I don't want to give a

subtask to like a sub agent if it's if

it's so big that it's like okay it's

going to start getting into like really

really high context zones like dumb zone

which like Dex calls it um from human

layer which I love and yeah so it's like

efficiently being able to take a problem

decompose it and then use like sub

agents as like compute sources to like

do those problems and like filter stuff

back to the main agent and like some of

it is just good model choice like for

example like we find that maybe the GPT

series like 5.4 for is exceptional at

like planning uh which is amazing and

like Gemini like I find is like really

really good at like multimodal stuff and

so actually so is they all are but like

Gemini is like really good at it and

like Flash is actually amazing bang for

a buck for like speed cost and

multimodal stuff like a lot of this is

just informed by like dog fooding and

evals like hey like we need to like test

these models and figure out what are

they good at so yeah I think I think

those are the three maybe roughly and

there's like way more obviously so it's

like like prompting

like systems around like verification

like self-improvement uh like via traces

or like via evals and then the last

thing is like kind of like orchestration

but maybe it's like context engineering

around problem decomposition

>> makes sense um you just mentioned about

uh 5.4 for for uh planning. So uh so uh

pretty much I think it uh not just a

black box but it is kind of a reasoning

sandwich where where I mean you

mentioned as well x high for planning

high for execution x high for

verification um like running only at x

high scored 53.9%

due to timeouts versus 63.6% at high. So

I mean that's counterative right? I mean

does more reasoning made it worse?

>> Yeah. So I think I think this is

basically touching on like the point

that I think about a bunch which is like

we need to like what we try to do is

basically like we're trying to design

like an agent system around like a task

that we need to solve right and like

that task has maybe like a bunch of

constraints like I think the one you're

talking about is maybe like the the some

of the terminal bench work that we were

doing and just trying to publish. So

yeah like for that use case we we had

like an artificial constraint which was

like we have a like a timebounded run

essentially like after this amount of

time like the sandbox just like exits

and like the run doesn't get scored or

like the run gets scored like wherever

we left the state of the sandbox and

yeah so I think maybe the takeaway from

that is less that like maybe like x high

reasoning all the way through like

wouldn't have been better. It actually

like does a great job. It just takes

like a really long time. So then it like

runs out of time to like complete the

task. But also it's like not compute

efficient and it's not like cost

efficient. Like it's awesome to like run

X high at everything all the time and

spend a bunch of token on like every

single problem. Like practically

speaking um you have to pay for the

tokens and like also like practically

speaking from like a user experience

like am I just going to wait for GPT 5.4

afford to just like think super hard all

the time or like can I use a smaller

model or like a cheaper model that I

like write really good instructions for

and it can just go do that task like

immediately then my user just like sort

of gets like a more you know like

latency reduced interaction. So it's

like yeah I think main takeaway is like

XH high actually for me is amazing and I

do a bunch of like planning in X high

when I'm like just coding but because

like when I'm in the loop I want like

feedback because like it's annoying if

I'm just like staring at a blank screen.

I use like high for a bunch of like in

the loop coding. So like X high planning

and then like high for execution. So but

yeah it just depends. It like totally

depends on like the work that we're

doing. I think that's like the main

thread that I think about.

>> Awesome. Okay. I mean yeah that makes

sense. saw and and I have seen that

people are using people are preferring

5.4 xi codeex over opus 4.6 six I mean

now seven has like mixed opinions I mean

anyways um so uh again like you said

about what about hardnesses and

everything and there was a potential a

lot of news about file system as well

like I can't give a count the number of

blogs I have number of Twitter articles

I have read about file system right and

even like in your anatomy post you said

that the file system is arguably the

most foundational harness primitive so I

mean it's a it's It's it's a strong

claim and um and previously obsidian co

also mentioned about everything just

about file system. So why the file

system and how does it kind of make it

really influential in in this harness

design and things around agent

engineering. What other tools?

>> I mean I'm like incredibly bullish on

file systems. I think like a ton of

people internally also are and like a

ton of people across industry like very

bullish on file systems. Like one of the

early decisions in like DB agent when we

were building it last year was basically

like using the file system and that was

more because we saw like two things. one

like how useful it actually is for

context management and like two agents

are just exceptional at using file

systems already right so it's like it's

kind of two things like the model is

already very very good at using this

tool so I don't have to coersse it a

bunch to get good at like using these

sort of like patterns and like now like

with newer models is probably even like

post trainer even more on getting good

at file system stuff so that's like

amazing the the other thing that's like

really amazing about file systems or

like basically the concept of a file

system. I I'll I'll maybe like

generalize it a little bit, which is

like I need some sort of like persistent

storage that my agent can use to both

like access information and then like

offload information. And like that's

maybe the higher level primitive like a

file system ends up being like a really

really easy way to do that. But like the

primitive is like the LLM the model

basically has like this computational

boundary that I put stuff into and like

I can take stuff out of essentially,

right? And like all the comput happens

here and the decision for like where to

store stuff and like how to access it

like file systems end up being fantastic

storage primitives to do that and like

the reason why I say like the concept of

a file system is like in in like lang

chain like in our libraries we have this

concept like virtual file systems where

it's like you expose file system like

storage essentially right so like the

operations that you would do on a file

system for example like ls for example

right or like you're like grapping over

that. It depends like what your

underlying storage system is. But can

you like use existing storage like for

example like S3 for example or like

Postgress, right? And then like what

does it look like to use that as storage

and then like put it over the

computational boundary so like the agent

can like search over this stuff and like

pull it into context.

Like agents are exceptional at doing

that. And the other thing is like

context management is so important

because like the context window is like

where all the computation actually

happens that we need some mechanism of

achieving that which is like why I'm so

bullish on file systems. It's both like

and then and then actually like maybe

one more thing I'll add is

>> now that we're doing a bunch more stuff

on multi- aent orchestration and like

multi- aent like collaboration sort of

stuff. So I think I said like a little

bit about decomposing like really big

problems into like sub problems, right?

But like where should all of that work

get stored for all of like the

decomposition that the sub agents do? So

like file systems actually also become

excellent like collaborations places. So

like sub agents can like write to

particular files and like main agent can

like read from there and like it doesn't

pollute like the main agent context

window a bunch. So it becomes like a

place where you just like write files

and like files are basically excellent

scratch pads or excellent like like

planning places or excellent like

persistent storage places like an agent

needs to come back to something and this

sort of like primitive that files encode

information really well like file

systems

offer like interfaces to like external

storage that already exists and like it

really helps with context management.

Like all of those things together I

think make it really really good for for

as like a harness tool for like an

agent. And I think a lot of harnesses

like like basically I think everyone is

like settled around file systems like

like it's uh it's not like too

controversial to say like I'm going to

give my agent a file system and like

that's a part of my harness you know

like people just sort of like oh yeah

that that makes sense. It's interesting

to know right I mean this is something

so basic something so fundamental is

kind of changed the whole trajectory of

the space in like 6 months and everyone

is kind of getting adapted to this thing

and on the same note you have uh you

have also mentioned about memory via

agents.mmd and and this is something you

kind of connect with you know like

injecting and start and you also call

this continual learning so I'm very

interesting to know about why do you

think So, and like is it really or is it

more like a persistent or consistent

notepad? So, what you really think about

this could be aligned to

>> I think like a a ton of a ton of like my

work recently has been around like this

just general idea of continual learning

basically. So like h how do I help my

agents which are producing a bunch of

data over time like I'm using let's

let's just take like my personal agent

like I'm using this one agent a ton over

time

>> and it's producing a ton of data which

is like traces essentially right and

then like all those traces like I'm

storing somewhere like we store them in

length you can put all your traces in

one place and how do I update the

definition of the agent in order to

learn from all of the data that it's

producing Right. So there's like maybe

two ways to really do that. And memory

is sort of a subpiece of continual

learning. Like continual learning like

overall to me is as I'm acting in the

world and as I'm like sort of like

producing data kind of like how we

humans do. Like I'm doing stuff in the

world and I'm like learning from the

feedback that I'm getting, right? Like I

ran and I tripped and I fell when I was

a kid and like this is a great trace

stored in my brain to say like please

like don't do that. Same thing for

agents. But the way that we actually

like update the like the agent knowledge

is like really different probably

because like we don't understand exactly

how like experiential memory that humans

experience like how does like my

experiential memory as a human get

encoded into my brain like I don't

exactly know how that process works and

we need to do that process essentially

for agents and like the agents

computation boundary is just it's

context window basically. So I need to

be able to like take learnings from the

past and I need to be able to like do

two things. One is um inject them into

the context window at the appropriate

time

>> so that when that scenario comes up, it

can like use that prior information to

like fix the thing. Like for example,

maybe this comes up in like user memory

for coding, right? It's like you're

doing a bunch of like coding with your

coding agent and then like you give it

it has that trace and like maybe you

like annotate that trace with human

feedback saying like hey like the way

that you did this or like you use this

library but like we never use that

library so like please like always use

this other library right and it's like

okay like great should that piece of

feedback and like context should that

always be in like my always on memory

right is that like just in my agents.mmd

that always gets like loaded in or is

this something that gets injected like

in real time into the agent like

contextually. This is like why I'm super

interested also in like search as a way

of doing this because like we're I think

it's like almost like unfathomable the

data scale that we're going to start

producing with agents. So like agents

run like all the time non-stop. they

produce like millions of tokens like

every few minutes and like that's a ton

of information that we need to like sift

through to figure out what's useful from

that and like what's not useful from

that. So like search is like a really

really big part of distilling a bunch of

trace knowledge into like nuggets or

like memories that I can actually

retrieve that are useful because like

tons of that trace will actually be

noise. So it's sort of this process of

like distilling

great data which is like trace data but

into nuggets that I can actually like

bring into context when I need to.

That's like one. And then the other one

is like really interesting for us is

instead of just selectively and

contextually pulling the right thing

over the like the context window

boundary for like computation to happen

over it. So like context engineering

like you can also just touch the

weights. So like we like lean in a bunch

into like open models and like I love

open models. I use like GLM5 a bunch

like a ton of the team does as well. And

that's like amazing as well. That's like

continual learning by using feedback

from traces and like distilling that

into data that you can do like RL on

essentially and like making that process

a lot easier. And both are really

interesting like we're leaning into both

and I think both will happen. So it's

actually not going to be like an or like

everything will be RL or like everything

will be like context entry. you totally

need both because there's like tons of

things that you don't want to RL or like

it just doesn't make sense to like

fact-based retrieval like you can like

include that data in there but it makes

more sense to do search in order to

retrieve some of that stuff. So it's

like yeah those are maybe the

interesting bits that we're sort of

leaning into like sort of

>> you just mentioned there are tons of

things which you don't want to RL so can

you mention what kind of arenas do you

think we should go for RL or we should

not like where there is like it is

constrained by compute resources or

anything

>> I'm like super bullish on if you're like

if you're a builder or a company

producing some sort of like data in

vertical and you want to like do two

things. One, make your model way better

at that task and like basically like fit

to your data, fit to your use case, then

also like make it like way faster and

like way cheaper. Like RL is something

like definitely like worth exploring

because fine-tuning has gotten like way

easier in the last whatever year. Like

there's actually like amazing companies

that will help you fine-tune if you like

bring the data, if you massage it

properly, like you store all your data

like Langmith and you can like pull it

down to do RL over it. Um,

in terms of things that you like should

RL on or you shouldn't RL on, I think

it's really really great if you have

some sort of like vertical that you want

to like make your model like really

really good at. I think we see a lot of

companies that have started, okay, like

I'm building this like model and it's

going to be really really good at search

and I'm going to expose that as like a

sub agent to like my main agent and like

this sub agent is going to rock at that

or it's like this this model we like

fine-tune on a bunch of our like

customer service data and like it's

really really good at that use case or

like finance data for example or like

even even yesterday um like OpenAI

released Rosalind right which is like

all about bio

That's like amazing, right? And that

also like sort of it it butts heads with

this whole idea that the general purpose

everything is just going to like kind of

like subsume everything, right? It's

like I'm going to have like one general

agent that's just going to like it's

going to be so good. It's just going to

get exactly what I'm saying. It's going

like solve the task. Like maybe in the

limit that is definitely maybe going to

be true, but to like today like we have

to build for today, you know? So like

today it's super helpful actually to

take the opposite view like curate a ton

of data and like pick a niche that you

really care about or like that your

customers care about and like build the

best data for that like build the best

harness for your model around that and

just like sort of rock at that task. And

I think like RL is amazing for imbuing

sort of like vertical specific skills

into an open model and you get it like

way cheaper like way faster and like

depending on the original like training

distribution of that task in like the

frontier labs like data mixture like

you're it's very likely that your

fine-tune model will be better than that

open model or sorry than that closed

model at that task as well because like

you have the data and you like

fine-tuned it and like maybe like where

you don't want to use RL4 is like I I

honestly think it's a really good idea

just to start with harness engineering

like or like just really good context

engineering

because it's so easy actually like

relative to RL that just like pick your

model like design like a really really

simple harness around it first like for

example we have like this abstraction

and lang chain called like create agent

which is just a react loop and then you

can like build a bunch of stuff on top

of that until like you don't need to

anymore or you can use like deep agents

out of the box if you want to and Yeah,

just like go and build and do maybe

start with harness engineering and like

maybe the other point was like

there's things that like things like

factbased retrieval like fact-based

retrieval is just it's just like maybe

more of a search problem like I just

want to find the thing and I want to

inject it into my context essentially.

So it's like yeah that might be like one

example where it's like hey like you can

RL this thing and maybe RL on the domain

but like the way that you put it over

the like boundary for computation the

context window is just find it

essentially via some search mechanism

>> you mentioned about search previously

right like you will be going for search

like essentially so uh there comes this

concept of context context ro so you

site the chroma research on how models

get words on on as context fills up

maybe compaction tool call offloading

skills as you know progressive

disclosure. So which of these has the

biggest impact in practice when when it

comes to context fraud and and what are

the what are the kind of potential um

practices you specifically use to avoid

these?

>> They they all matter actually and I I

think it it sort of depends on like the

design that you're going for out of the

box, right? So I think like maybe maybe

like a good recipe essentially is that

like we start building the agent with

like a goal in mind like I want the

agent to do this thing but like really

really focus on like context rot because

after you pass like some sort of like

context threshold it gets like just like

really dumb and like like you said we

have like levers to fight against that

which is I can use like sub agents to

decompose the problem into like more

manageable chunks so I don't pollute my

main context window like that's like

amazing but like basically what what

like predicates that is that I can

actually efficiently decompose the

problem, right? So it's like maybe

that's like instructions that I give in

the system prompt to the agent of saying

like this is how you like decompose a

problem into like these tasks and like

if it's like a task specific agent then

you probably already have a bunch of

like human priors for how to go tackle

the problem. Like for example, for

coding agents, the way that we decompose

a problem is like you have agents that

do like sub agents do like codebased

search essentially and they like do that

separately and they pull in the

important information into like the main

agent to do some of that stuff and like

maybe there's like a web search agent as

well like

>> has to go like pull external information

and like find that and prepare it for

the agent. So it's like yeah basically

like working backwards from like I need

to avoid context rot like one way to do

that is like sub aents like is my

problem amendable to like sub aents like

if it is fantastic another way to do

that is like and these are often in

conjunction like we we like lang like

our docs we publish like a bunch of

stuff on like multi- aent docs as well

and skills are kind of related to that

which is skills to me they basically

kind of like encode knowledge and

workflows like skills are awesome

because everyone before skills like

hated writing good docs if that makes

sense. Like everyone was just like so

lazy

>> and they were like I'm just going to

like tell the model like some sort of

like random stuff like kind of like hand

wavy and it'll just get it. But like for

some reason like skills came out and

like because maybe skills are like

sharable and like other people like see

the skills like everyone writes like

very very good like workflow

descriptions in skills and like the

agent sort of like sees the skill

content and then it executes the

workflow and that's like amazing because

>> I basically get like a very small

snippet of like when to use this skill

and like I avoid all the context rot and

like when necessary we like pull in the

right context from the skill into there.

Like the the tricky thing with skills is

always like

basically knowing when to trigger them.

And that again comes down to like

instruction following which is like we

have some skills evals as well where

like we'll have scenarios and then we'll

sort of have like the skills that we

want to have like triggered basically

and then like we we have evals where

it's like we we only want that skill to

be triggered because like like let's say

it like triggers the wrong skill first

and then like eventually it does a bunch

of like stuff and then it figures out

like oh actually I need to do this skill

like that's bad because you wasted a

bunch of tokens essentially. So I think

eval help a bunch with context rot which

is one does the problem succeed at the

end like that's a really big part of

evals and the other one is sort of like

fine grained metrics on the evals which

is like how long did it take like how

many tokens did it take what was the

overall cost right and then like reading

the trajectory and then seeing like in

my effort to reduce like context rot by

doing like sub agent routing or like

triggering the right skills is that

working and then there's like maybe also

like determine ministic stuff which is

good. So like tool call offloading. So

like this happens a bunch with like bash

calls like you have you you you run the

shell and uh it's just like a mess. So

you get this gigantic like tool like

this output string and you can just pipe

that into context or you can just take

like the head and the tail and you pipe

that into context because that's usually

the important bits and then you tell the

model that like the rest of this string

lives in this file over here if you can

if you want to access it and then you

can go and do that. So, it's basically

like doing a bunch of stuff on the

model's behalf to really protect that

like incredibly precious artifact, which

is our context window. And like we I we

just think like really hard about like

if something doesn't need to go in here,

like really like don't put it in there.

But if something does need to go in

here, like do our very best to like

spend compute on like search or like

really good instructions to make sure it

gets in there.

>> Makes sense.

Interesting. I mean that that actually

makes a lot of sense. Um I'm curious

about so for people who may who may not

know the space well. So there's been

like open claw boom. I mean I I just saw

on Twitter it is kind of declining as

well. So Hermes has been getting as much

attention as open claw right which which

is coming out of news research. So how

does

deep agents differ from both of them? It

would be useful to explain this from

first principles for both technical and

nontechnical listeners since we are

going to spend a lot of time talk about

hardness in this conversation.

>> Both amazingly sick projects like

openclaw amazing like what Peter did

there and then also like what the new

guys are doing with Hermes is like so

cool. Yeah. I think like the main way

that I think about it is like you have

this like claw architecture right looks

like a little bit different from like

claw to claw but like overarching

architecture of like I deploy this

somewhere there's some sort of like

messaging

>> it's like live talk to it back and forth

there's like a heartbeat that triggers

like over and over again that has like

some sort of like memory primitives in

there so it's like it's basically like a

very opinionated

harness for the use case that is like my

personal agent. So I think like a claw

is like really the first it's the first

like really mainstream personal agent

like maybe like besides chatbt like

chachi is like it didn't like really

feel like a personal agent like it had

like memory and stuff like people like

message their claws like all day like

maybe people do that with chatbt too but

it's like the architecture of the

harness behind like the claw that makes

it like feel really personal because of

all the things they put behind it around

like the integrations like like what's

happening and telegram and those types

of things and like the memory that gets

updated. Like the big thing is honestly

like I like the heartbeat thing a lot. I

think like that doesn't get enough hype.

It's like very ingenious to like wake it

up on some cadence to like do things for

example crrons and things like that. So

the way I think about it like high level

is a claw is an amazing choice of an

opinionated harness for like a personal

agent essentially. And that's like an

awesome choice that like they make. And

I think we should have like a lot more

of these like people should like build

their own or like people should use them

more and see if they like them. And then

maybe like going back to like the

primitives, I think like you can build

tons of agents that are not claws that

like completely solve like your task

like really really well. And that's

basically like how I view like maybe

like Langchain's create agent or like

deep agents or like all of the other

great companies that are like building

harness primitives which is your

probably your task does not require like

a claw like most most likely like it's

awesome like you should have a claw in

your life but like if you're doing

something else like you don't need a

claw. So like actually what you need is

amazing instructions, amazing context

engineering, like amazing choice of like

what models you're going to use to hit

like the paro frontier of like Perf cost

and latency and like you can start from

like a simple harness and you can

assemble a harness around like that

model or like models to like build that

thing essentially. And I think like claw

is like one instantiation of an

opinionated harness for like personal

agents basically and it's like awesome.

And I know like people use claws for

like other things as well. So like I

think claws if you like edit the harness

around them like the base harness and

you like make them I don't know if you

like change them for like another task

of research that's like awesome too. But

I think like that whole process of like

taking a task and like you have a

harness that like wraps a model or like

models and like you sort of like direct

it towards a goal. That's basically I

think the goal of like laying chains

like create agent and like deep agents

which is we have like some opinions in

there to get you started but like really

we want to help you build the best agent

like for your tasks. Like that might be

us giving you all the tooling. That

might be like me and like the rest of

the team like blogging about like actual

use cases and like sharing our evals and

like just publishing results. But yeah,

basically like customize a base harness

to make it like really really good at a

task and like a claw is a like

phenomenal example of

>> makes sense.

What what you really see the future of

it? I mean I mean let's say down the

line what the next um what what it would

look like after let's say five

iterations of it or what you really see

the future of in in a year or so let's

say

>> I mean honestly in a year and six months

like everything's going to change

obviously no I'm just kidding like it's

it's hard to say like I'm like in the

short term very very bullish on

basically helping people build like open

or like us providing open infrastructure

to help other people build agents that

are like amazing for their task. And I

think that is not going away in the near

to medium term at all. In fact, I think

it's going to go in the complete

opposite direction, which is like

everyone is going to start basically

taking their tasks and they're going to

either do like harness engineering

around those tasks, which is largely

like very good context engineering, like

very good like prompts and very good

tools and like very good skills. They're

going to do all of that around like some

sort of task they care about basically.

And I think open harness engineering is

a big part of that. And I think like

open models are also like a really big

part of that which is we're going to see

like a big growth of I'm going to take

like Kimmy, I'm going to take like GLM5

and I have this data and like a big

future is I'm just going to like

fine-tune that model on my data and I'm

going to make it really good. I'm going

to just keep doing that over and over

again and I'm going to compare how that

does to like a Frontier model and I'm

going to make the trade-off between like

is it better, is it like just as good,

what's like the cost, what's like the

latency tradeoff and then like maybe

like a little bit more like longish term

from that which is like it would be

awesome if we got like some sort of like

AGI model that just did everything. I

would love that. Like so then I can like

totally stop talking about like

harnesses and like evals and I can just

like enjoy the model. But it still

really does help to specify like the

intelligence that we want that model to

like act on in a particular situation.

And like I still think even in like the

medium to long term, it's going to be

super helpful for humans to get really

really good at both describing the thing

that they want and not just like hand

wavy like writing like kind of how we

like write really detailed prompts like

getting comfortable with taking the

thing I want and like putting it into

like language basically. And then the

other thing is we're still going to want

to like verify the work that agents are

doing in in some way. I hope like

autonomous verification systems get a

lot better, but like they're not going

to be perfect and we're still going to

want to be able to say like when an

agent is doing good versus like when an

agent is doing bad and that can become

like part of the feedback signal and I

still think that's going to exist like

for for a little bit and like that's

that's like not a bad thing. That's like

totally fine like we can still work on

that.

>> Makes sense. Um you know dude there are

a lot of I mean bunch of companies I

mean interestingly everyone who is

working on frontier are coming out of

their own harness own their own agent.

So so recently RAM basically built their

own harness right. So have you seen what

they put out? So and my other question

is like you yourself um have used open

code. So it seems like um enterprises

building custom harnesses puts real

pressure on competitors. So I'm guessing

a um release like that forces companies

like SLA which also got recently big fun

who competed to RAM and others to build

their own thing too. So what do you make

of this trend?

>> Yeah, I mean like the ramp is amazing

obviously like they put out such fire

blogs. I like ramp lab stuff. I think

like the the overall trend of like

building your own harness or like or

basically like building your own agent

that's custom for your task is like

fantastic. Like I I think more teams

should basically devote time towards

like investing maybe like in the process

of both like helping their teams build

agents, right? That doesn't just mean

like coders. That means like everyone

like the people who are doing go to

market like marketing, sales, all those

people can like benefit in some way from

agents. They just need like help doing

that basically. And I think it's great

that a company basically like picks a

problem and they're like we're going to

solve that by building the best harness

and that means like the best context

edge like the best verification the best

tool stack also a big part of it and

like we work on a lot of the stuff at

like lang is building the correct or

building like really easy to use systems

for taking the trace data and then like

improving the agent because I think

there's a lot of stuff around

improvement loops which is our first

pass at the agent isn't amazing. So like

this comp these companies are like okay

I'm going to pick a task I care about

and like my first version is going to be

like kind of mid totally fine but then

I'm going to get the data from somewhere

and I'm going to like make it better

over time by just like spending a ton of

time on it or like maybe spending a

bunch of like compute on it to like

understand the data and like improve the

prompts like fix the edge cases like

improve all the errors right and it's

just like I still think like we we will

have tons of vertical companies because

today like someone has to do the work

like someone has to like invest in doing

that like someone has to do like sales

around that right it's not just going to

like happen by itself and I think like

more tooling around that and like more

yeah just more like research that helps

people do that that's like a good thing

like doing the open is like an even

better thing so it's like more

>> I think it's also very ambitious to do

you know you you're you're already at

Frontier and why to depend on someone

like let I mean if you want to be at

frontier you have to build something

like what what other people at are

working on. Interesting. So, um beforeh

going to the other segment of the

podcast, let's have some quickfire

chats. Uh so, so there is a meme you

liked from Mintly Fly Slack. Would be

awesome if you can share screen and

share that. Yeah, please. Then I'll go

ahead.

>> Let me

let me get that off. Um

I love this guy. Let me share. Dude,

this guy is so funny. A dude, I love

this guy so much. Dude, this guy this I

don't know like where this came from or

like who can I turn on volume?

>> So then I mean you like this like this

from mental slack channel that

apparently went viral across I think

startups. So what is it and what it made

it so hard? What's

>> so I have I have no idea. I think it's

Nick. It's Nick from Mintify like

tweeted it one day who's funny like I

was just like this is amazing. So I I

just sent it to all my friends um just

like randomly. I think it was our like

uh like soccer chat. I'm like something

happened with like Arsenal or something.

I I sent this to like my friends because

they I think they lost. Yeah. And like

now we have this in our Slack as well.

Like someone made it into like a gift

and like whenever maybe something goes

wrong like we just sort of throw this

guy like I don't know what it is or who

made it, but like I love this guy. I use

it all the time.

>> Awesome. Um my next question is um most

underrated harness feature that nobody

talks about. most underrated.

It's a good question because like I feel

like if it's underrated, we should be

talking about it.

>> Yeah, exactly.

>> Okay. Okay. Okay. I think like one thing

that we use a bunch is like this idea of

like we call it middleware but like

hooks just generally. So like for for a

lot of teams it's like super useful to

inject sort of like deterministic

actions like basically like do

deterministic code execution like

somewhere in the harness. And I think

that's like super underrated maybe

because it requires like sort of like

custom logic. It's not just like you

think of a tool and you just sort of

like add it. But yeah, I think hooks

that sort of like control bad model

behavior are like really really helpful

or like not just bad model behavior like

help the model like do things. So for

example like triggering excuse me

triggering like self-verification and I

think people should like build more

hooks to control their models. Makes

sense. Um interesting. So we have

something which people should talk

about. Interesting. The model that

surprised you most in agent workloads

this year

>> in both I mean we can go in both ways

which was like something which you were

not expecting and it comes out really

better and something which you kind of

not expecting and it was like it comes

out.

>> Yeah. So, I'm like so impressed by open

models generally as like actually ways

that I get work done. And like I think

like it's it always like feels really

good to talk about open models, but like

you sort of like love the idea of open

models, but then like you don't use

them. Like that's like that's not good.

But actually like the open models that

have come out this year are like

amazing. So like the GLM series is like

fantastic and like it is actually a good

agentic coding partner. It's like very

fast and it does amazing work. So like

maybe at the start of this year like

last year I don't think I would have

expected my GLM f like my GLM use to be

so high and there's other models too

like um like the Ry team who you had on

like they're they're amazing. Miniax is

one that we actually like eval on a

bunch and like these are all amazing.

So, like open models have surprised me.

Like I was hoping it would happen, but

it did happen and that's awesome and

like we should invest a bunch more in

that and like I hope like I hope like

teams actually like think about using

them in like their actual workloads

because they're amazing. Yeah, that

surprised me but in a in a good way. I

was like super happy and like it's only

going to get better and that's like

really really good and it's like way

cheaper and faster.

>> Awesome. Um okay, this one is lost. one

thing you would change about how the

industry builds agents right now. It can

be any common practice or something like

that.

>> How should they change? I think

basically like this whole thing that

we've been talking about right now is

like I would love if like that was like

easier for people to do or like more

people like did it which is basically

like maybe like work backwards from a

task and like a goal that you really

want and then like the whole point to me

is just like build a system like for

your team or like for yourself and like

for your agent to like make it better

over time. Like maybe like I'm saying

that because I'm like we're thinking a

lot about continual learning. So this is

like both the agent design which is like

prompts tools like the whole harness

thing like the verification loops like

all this sort of stuff and then also

it's like sort of the infrastructure

around it for doing like

self-improvement. So this is like the

unsexy stuff, but I think the stuff that

like really matters, which is okay like

are you like is tracing on? Like are you

like putting your traces somewhere

basically like are you using your traces

to like mine errors like via monitoring

basically? like lang supports that and

like we think about that a bunch which

is like trace came in like how do I

figure out if something happened and

like am I making eval from that right

and then like am I am I like reading the

evals basically so it's sort of like the

systems approach around like building an

agent and like making it better I think

teams are doing that that's amazing but

like it's awesome and I think teams

should team should try to do that

>> makes sense um also on the same note

there was this um recent paper called

meta harness and DDR also posted about

it a lot of people are working on auto

research and this field adjacent if not

a version of auto research itself then

you have also things like you know post

train bench where a hardness is used to

post train models so if those two

directions start merging so better

harnesses improving post training and

meta harness improving the hardness loop

itself I mean that feels pretty

explosive pretty interesting how do you

think about that convergence

>> I I think it's super exciting like I

love teams that are like

productionalized ing like auto research

and like doing so like we we have I

think we did like something around like

harness opt maybe like a couple months

ago and there were definitely like some

issues that I saw back then and I think

we still have like some of the issues

but like now like a lot more teams like

putting a lot more effort into it. So

it's like I think this is amazing that

and also maybe I'll just like this is my

take as well. So, like I have like we

put up a bunch of blogs and I think like

there's there's like algorithms that we

need to discover to make like agents and

harnesses better using some sort of like

grounding signal. And that's basically

like auto research is like I have a

grounding signal and I hill climb that

signal and like I update my harness and

like meta harnesses that like we have

one like better harness and like we tons

of good people have like work around

this which is amazing. And basically

like I view like eval as such an

important part of this like feedback

loop because like eval are basically how

we like ground our like auto research

loops over time and it's not just like

ground like in the moment like if I run

auto research like later like a two

weeks later I still have that same like

grounding mechanism and maybe hot take

but I think like you can almost try to

define an agent via a set of like evals

that sort of serve as not just a spec,

but a spec that you can like verify and

like ground. You can do it via fitting

to like eval.

And then you basically have a fitting

algorithm, right? And the fitting

algorithm can be like meta meta harness

or like better harness or like any of

these. And that fitting algorithm is

basically run on evals like reflect on

traces and like update harness or like

prepare data to do like RL on it. And I

think like we're in such early early

innings of this self-improvement loop

basically and I'm I'm super excited

about it. I think it's like really

really cool. There's like stuff to work

out around overfitting and stuff but

like that will happen and like people

will use this a bunch more.

>> Awesome. Makes sense. Um

>> I'm pretty much looking forward to as

we're approaching to the next um I mean

last segment of the podcast we have some

questions around environments harnesses.

pretty much harnesses we have covered

but eval and benchmarks around

benchmarks. So, so how do harness fit

into this broader idea of simulation as

a service? So there are companies whose

whole business is simulating work

categories, decisions, operating

environments. If better harnesses lead

to better simulations, so where does the

open-source side go and do you think

langen will eventually release an open

source simulation?

>> Yeah. So like I think these these two

things are super related. So like thing

that we wrote about before there's like

like evals and like environments like

they're not the same but they sort of

like rhyme with harnesses as well. So

it's like basically like the like the

main idea is like I need like some place

for my agent to do work that sort of

like reflects actual work that's going

to be doing like in the real world

basically right so it's like I'm going

to build an environment like there's

tons of like awesome environment

startups that are doing that and like

running the agent in them so it can

produce like a good feedback signal so I

can like train on basically that's like

amazing. I think like even a big part of

like evals are going to start looking

like environments because like when when

we first started trying to like eval

it was really simple. It was like chat

completions evals, right? It was like

I'm going to give you like a really

simple like input prompt and I'm going

to have like a number or like a

structured output at the end of it. I'm

just going to like map the keys, right?

I'm going to be like, "Hey, did you like

did you get them all right?" But as

agents are doing much more like

complicated work and like much more like

long horizon work actually like the

thing I want to eval is like a task

essentially and like the the best way to

maybe do that is to just like build an

environment and just like drop my agent

into the environment and like maybe like

what we do right because we actually do

this like we basically use Harbor right

and like Harbor those guys are awesome

like the the terminal bench guys

so like we'll pick our eval like it maps

to some sort of like hardware config and

then like we run the eval in the

environment that we built. Then like all

of the traces like go into Lenmith and

then we like read them and we look at we

like segment them based on like the

rubric like how much did it pass, how

much did it fail like how long did it

take and then we try to like improve the

agent and I think that process of like

building the environment and like you

asked about simulations like we think I

think about this a bunch which is like

what I really want to happen is the like

the company that we're building or like

the app or product that I'm building

like I want my agent to be able to like

test itself in that exact environment.

So it can figure out like when stuff

goes wrong essentially and then I can

like fix it, right? And like that's

basically the whole point of eval which

is like

>> they're sort of like a proxy for what

happens in production and like as I fit

to my evals I'm kind of imbuing like

behavior into the agent to make it pass.

The whole goal of evals is like to make

them pass, right? Like and like a lot of

our evals fail because like maybe the

models just aren't smart enough yet.

like eventually they will pass and like

then what I've done is like I've taken

that information from that eval and I've

sort of like transfer learned it into

like some sort of agent whether it's

like the weights or like the harness or

something and yeah I'm bullish on both

I'm bullish on like eval as a mechanism

of doing like agent improvement and also

bullish on like more eval looking like

environments basically instead of like

just like input output

>> pretty interesting you know this uh

terminal bench 2.0 Sweet bench pinch

bench. So the benchmark landscape of for

agents is growing fast but you um

explicitly say in your opinionated

agents post where I mean you said test

on real world users for your product

don't trust benchmarks your user has

never heard of terminal benchtop please

don't introduce it to them right

>> so so so so which benchmarks do you

actually trust and which ones are most

performance theater I mean I mean what

what's the general landscape where one

should actually think

>> I was definitely a bit hyperbolic like

like don't introduce anyone to Turbo

Veg. Like I love Turbo Veg. Like those

guys are awesome. But I think I think

the general point actually does stand

though, which is

like like eval to me are basically like

again they're like a mechanism of like

evals and benchmarks. They're basically

like a mechanism that like proxies

behavior that I want my agent to

actually have like via this like thing

which like roughly measures that, right?

So it's like I'm trying to measure like

long horizon like problem solving. Like

can I do that with like a really hard

terminal bench task? Like kind of maybe.

But like if my actual like app has

nothing to do with that, then like me

passing that like terminal bench task

doesn't map well into like my like long

horizon problem solving like the bio

domain, right? So it's like there's sort

of like rough proxy signals that measure

like so like in at like Langshin we have

like axes that we try to measure on. So

like every eval we tag to like an

access. So it's like retrieval, it's

like problem solving, it's like

planning, it's like tool use for example

like we we like try to tag every eval

we do that we tag every eval like one or

multiple axes, right? But I think it's

useful to use benchmarks as like a

general like vibe like a guidance. Like

you should definitely read the traces

from benchmarks. We like I like spend

tons of time every day just like reading

the traces from pre-built benchmarks.

But I think really the thing that helps

teams is

using their trace data to build evals

for themselves that actually like map

onto their customer use case that maybe

like no existing benchmark like does a

really good job of. And I think like

it's kind of like a moat if you want to

call it but it's like it's just a really

good way of like building a better agent

product which is there's awesome people

building awesome benchmarks. None of

those benchmarks map exactly onto like

what my agent needs to do. So like I can

like use those to roughly measure

problem solving ability, but like really

the best way to measure problem solving

ability is just to like get a

representative set of like evals and

tasks like my own bench and just like

use those and like that's going to vary

from like person to person like product

to product like feature to feature. So

>> makes sense.

>> Yeah. Yeah. Yeah.

>> Yeah. What's your opinion on computer

use stuff? Because this is u this is

something very subject to people like

the current approach is not good. You

can't really go in the screenshot way.

You really can't use MPI MCP or API way.

You have to bullish. You have to scale

GUI stuff. So what do you think about it

and about it scaling part of

>> dude? We were just talking about this

today actually like how much should we

do like more examples on computer use?

Like I'm like very fascinated by

computer use. I think it's like super

interesting. I think like there's maybe

like two things. One is there is still

definitely a visual perception problem

like that we like we've known for a

while like like fine grain details is

not it's not like amazing at that maybe

it's like less of a limitation now like

some of these models are like better at

computer use I don't know I don't have

like a great opinion on which way of

doing computer use is going to win like

the hybrid like pulling down like the

actual like webpage content and like

clicking versus like how much do you use

screenshots I would be like very happy

If like everything was just like just

worked with vision because that would

mean that we did we have made like a

step change in like visual perception

and like visual reasoning over

screenshots and like doing or sorry like

yeah visual reasoning over like

>> images basically.

>> Yeah, I don't know if it'll happen. Um

but I'm like for the applications of

computer use I think they're awesome and

like we should do we should do like more

stuff around those but I don't know. I

just haven't played with it as much.

>> I mean yeah awesome. Makes sense. So I

mean do you think there is some secret

sauce something which can be scaled to

scale more long horizon task about in

your experiment experience what what is

something which is blocking

uh because I think a lot of companies

are being forming up involvements for

long horizon task these days and been

selling to enterprises and frontier labs

now I mean what do you think of the

space about scaling

>> I think

>> I think like there's a lot of like good

work that a I think a lot of companies

have like good agents in like medium

horizon tasks like for example like we

like we have like a background coding

agent that can go like do things over

like hours like a few hours basically

right it's like they're all coding

related tasks like it's easy to like

pick those and like scale those and I

think yesterday there was like really

good work from like the proximal team

for like frontier suite which are like

hey like these are like 20our tasks

basically and like we're going to go and

run on

I think like one thing that is still

like really really tricky for models and

I think like in the near term what's

will happen is like hopefully like

models get like post-trained better on

this but like we will still have to

build a bunch of like harness

infrastructure around it which like

hopefully falls away is one like

decomposing a really difficult problem

into like subpieces

and then doing like verification of the

intermediate steps. I think like that is

like a really really good general

purpose recipe that we can use to like

keep doing like longer and longer

horizon tasks because like basically

like all a long horizon task really is

is like I'm going to do I'm going to get

this like really hard task. I'm just

going to do like a bunch of like little

sub pieces like over and over and over

again. And I need to make sure that like

I don't mess up any of the sub pieces or

like if I do mess up I need to like go

back and fix those

like the key thing is like figuring out

like when you messed up that's hard. So

so we we need better like

self-verification systems there that

might be like self bootstrapping like

testing for example

>> and like the other thing we need to do

is like teach systems how to like

decompose problems into like sub agents.

I think like there's really cool stuff

around RLMs around this.

They're like I still find them like a

little bit tricky to get working, but

like the ideas behind them like amazing

basically like externalize context as

like an object and then like sort of

like search over that and like decompose

problems like that for like really

really long horizon tasks. I don't know

it doesn't work amazing right now but

like the general strategy of like verify

and then like decompose like iteratively

I think that's like a good path forward.

like we're we're spending time there as

well. I'm sure like a bunch of other

people are well.

>> Awesome.

Great. Um I think we are pretty much uh

to the end of the part and um so what is

uh what is something which you are most

excited about to happen in let's say

again in 6 months or year because again

like we pretty much don't know but you

really want to see to yeah to happen.

>> Two things like one I'm super excited

for the World Cup. So like World Cup is

happening like here it's happening in

Philly. So, I'm like super stoked for

that. But besides that, I think like the

thing I'm like super stoked about is

we're we're like just starting to get

the first sparks of these like

self-improvement loops from data that's

generated from agents. And like we're

pushing like a ton on this like in the

last like couple months like we put our

like first like research around this.

There's other good teams doing this. But

I think like this is like such an

amazing on-ramp for us for like all

teams to self-improve like all of their

systems by doing like very very good

like data engineering like looking at

all of their trace data like mining it

for errors and like bootstrapping self

like probably to start they're going to

be like semi-autonomous self-improvement

loops like like humans will need to be

in it but the systems will get better

and better and I think the the flow of

build agent

use an environment, generate data from

it, and then like mine the data, point a

lot of compute at the trace data to

derive like eval and to derive training

data and then like use that to like

improve the agent. Just keep doing that

loop. That is like super exciting to me.

And it it's it already works actually

like every like people like we're doing

it like people are already doing it. It

like works. Customers are doing it. It's

awesome.

But it will only get better, I think,

with like better models and like we're

going to build everyone's going to build

better systems around some of this

stuff. So, yeah, I'm stoked in six

months. Like, I can't even imagine like

how good this loop is going to be. It's

going to be amazing.

>> Likewise. Totally. What's the next blog

coming?

>> Next blog. Um, okay. I'm supposed to

write one over this like weekend. Yeah.

Hopefully like next week. Yeah. Oh,

maybe like one thing that's cool I like

lang chain is like because we talked in

the beginning I actually think blogs are

like fantastic like artifacts like work

backwards from so it's like but your

team does a bunch of like amazing work

and like you should like totally share

that work so you can like kind of like

pick like a blog it's like I want to

write a blog about this and it's like

okay like what's all the work I have to

do to make sure that that blog doesn't

like suck basically. Yeah. Yeah.

>> That's great.

>> Yeah. Yeah. I think there's one I'm

thinking about a bunch which is like um

it's less like like agent engineering

stuff but more just like how much we've

like unbundled like agents. I think

there's been like a huge like unbundling

of agents uh into like pick a base

harness and like pick your skills like

pick your tools like um design your

agent like design the models and it's

not just like one monolithic system like

you totally don't have to get locked

into anything like you have the choice

to build like bespoke tooling for

yourself like for your company and like

the unbundling is awesome and like I

think like people are doing cool stuff

around that so hopefully like I'll like

riff on something about that or

something or just whatever I don't

Great. Okay. We um last question to you.

Um so imagine so um so the world is the

technology is changing

by an order of magnitude every week. we

all can see uh what advice would you

give to someone who is just starting out

of college who is someone 20 20 21 year

old because because things are not same

as it has been like I can say it's been

like like couple of years ago it's not

the same the world is changing so fast

and and it's sad to see that lot of

people are actually I mean don't even

care about what is really happening

right so even like if someone is

starting out college so what should they

really look forward to? I mean to to be

at frontier and to actually scale on

things to actually learn and be at good

places. So what's your opinion over

that?

>> I don't know how amazing advice I can

give on this honestly but like maybe

like some like general thoughts of like

what I was thinking when I was like

finishing like PhD and stuff and also

like there's like so many sick like kids

who are just like graduating undergrad

already like that I see on Twitter doing

great work. I think like there's there's

a couple like common threads which are

really cool which is basically like you

just sort of like pick something you're

like kind of interested in and you just

use like AI to help you learn that and

you just like kind of like rabbit hole

like really deep into that one thing.

And I think like that's probably like

really really useful because you can

kind of maybe use AI to become like top

maybe like 10% or like 5% of the world

if you like care enough and like the

problem is not like super crazy. And I

think like that's like really good. And

the other thing is like I think like

it's awesome when people just like post

their thoughts like online. And um I was

saying like it helped me like meet a lot

of like cool people. I see like awesome

like posts on X and like I love

interacting with them, but I think it's

basically just like it's kind of scary

to maybe like put your ideas like online

like dude I'm gonna get like roasted

like first I'm going to get roasted like

by my friends who are like dude why is

he posting so much on like Twitter about

like AI but it's totally fine like you

kind of like get over it but like it's

just like really good to like sort of

share your ideas because it helps you

like other people like challenge you and

then like you realize like oh like that

idea was dumb or maybe that idea was

like really good like resonates with

people and like the only way maybe for

like other people to like really help

you is if they like see your work or

they see your thoughts and then like I

think there's so many people who are

like willing to help. So just like maybe

like pick something just like grind on

it just like post about it basically.

And I feel like if you do that enough

times then something good will hopefully

happen or like you'll have learned

something which is also like really good

>> dude. Um this is so honest and I can

totally relate with both of your points

and basically this is something which I

have experienced again like because

there are so many trajectories so many

arenas opening as AI is evolving to

learn to to actually u make your make

you context aware about things I mean it

can be anything it can be posting side

of things pre-training inference

engineering environments data a lot I

mean you can't really keep it up about

things so again as you said use AI use

your knowledge sources and like read

good blogs, references, hack on,

experiment on and this is something and

that is the reason I mean even good

professors lot of colleges are not

actually wor about things. So I think

this is the best time to learn and

actually dig on things and and I think

there are wide arenas where one can

master one thing right because everyone

needs master of something and get into

places and let's I mean it can be

anything it can be even hiring as well

if you're really good at it so you can

make it to the places of course and as

you said about posting about stuff dude

I mean this is so underrated I mean if

you are really good poster if if you if

you can really uh kind of um convey your

thoughts well

amazing opportunities can open up and

this has been happening for me and I

have seen a lot of amazing people been

to places just by like I've I've

interviewed bunch of people I can give

example of kalome he's he's 19 he just