Harrison Chase(@hwchase17)

this is great deepdive into what a managed harness looks like! some quick thoughts while reading: -...

Harrison Chase (@hwchase17) — Tue, 20 Oct 2026 21:01:58 GMT

this is great deepdive into what a managed harness looks like! some quick thoughts while reading:

- many diff types of agents built on it. doesnt surprise me - a lot of agents look very similar inside these days

- why not use off the shelf? (1) model and provider agnostic), (2) co-build with product teams, (3) own observability stack. I very much agree with points 1 and 3!!! and is something we obviously emphasize at langchain

- "The simplest agent harnesses are truly just a while loop, but when you want to support a wide array of agentic products in production, the responsibility of the harness or framework expands quite a bit" - +100 to this, is why were investing in deepagents so much

- "Using the “sandbox-as-a-tool” paradigm" - the massive trend seems to be going this way (is how deepagents is architected)

- "Supporting non-blocking async agents as the default". we are working on this right now for deepagents! agree is the future

- "The Connectors Service is the standard interface through which agents at Ramp interact with data sources. It centralizes authentication, credential management, access controls, and tool execution behind one API. This allows us to build and maintain integrations once, and reuse them across products." Also very important! and something we're thinking a bunch about for managed deepagents

Kiryl@kir_lev

We at Ramp tried a bunch of agent harnesses and ended up building our own. Cloud-native, model-agnostic, tailored to our infrastructure and product.

And we published how we built it: builders.ramp.com/post/building-…

🔗 View Quoted Tweet

💬9🔄3❤️28👀2846📊13

⚡ Powered by xgo.ing

love to see this integration

Harrison Chase (@hwchase17) — Mon, 19 Oct 2026 13:50:42 GMT

love to see this integration

Emil@emilsnotes

introducing @LangChain × kapa

your langchain agents can now search all your company knowledge more accurately, 3x faster and 8x cheaper than frontier models

we synch docs, github, notion, slack, zendesk and more into a single index (with vectors + file system)

your agent gets two tools: search everything, and read the full doc

it takes 2 min for your agent to set up (+ free to try)

just paste this into your agent:

"set up a knowledge base for my agent using docs.kapa.ai/getting-starte… and pip install langchain-kapa-ai to use official plugin"

keep making agents go brr @hwchase17 + team!

🔗 View Quoted Tweet

💬23🔄2❤️47👀7171📊21

⚡ Powered by xgo.ing

Most of the interesting agent harness work is happening at intersection of harness and infra: 1/ dr...

Harrison Chase (@hwchase17) — Thu, 15 Oct 2026 19:23:20 GMT

Most of the interesting agent harness work is happening at intersection of harness and infra:

1/ dreaming 
2/ async subagents
3/ heartbeats
4/ letting agent schedule reminders

Part of reason im so bullish on managed agents platforms!

All of these in managed deep agents!

💬167🔄36❤️487👀25233📊241

⚡ Powered by xgo.ing

Coding agents (Claude code, codex) -> personal agents (instinct, muse) -> companyOS (???)

Harrison Chase (@hwchase17) — Thu, 15 Oct 2026 11:37:22 GMT

Coding agents (Claude code, codex) -> personal agents (instinct, muse) -> companyOS (???)

Yohei@yoheinakajima

after switching our approach to building a "FirmOS", I'm increasingly confident that the future of agent collaboration within organizational settings is going to be a "company operating system" or CompanyOS, designed for agent and human collaboration

instead of a "company brain" that holds everything, the OS is for storing rules, processes, decisions made, which tools to use and how. it can mirror some data, but the source of truth is often best held in your existing tools (CRM, etc.). it's continuously updated and used by your agents

i think of this as a "flipped playbook" where instead of writing down the company rules first and following them (which is kind of like waterfall development), you're constantly updating the playbook as you decide or learn new things day-to-day running the business

🔗 View Quoted Tweet

💬151🔄12❤️153👀17722📊126

⚡ Powered by xgo.ing

router lives in the harness, not a gateway, bc the harness knows the task blog: https://t.co/k9b306...

Harrison Chase (@hwchase17) — Tue, 13 Oct 2026 16:16:08 GMT

router lives in the harness, not a gateway, bc the harness knows the task

blog: langchain.com/blog/how-to-bu…x.com/neil_xbt/statu…

NeilXbt@neil_xbt

LangChain gave its coding agent three models and a router.

Open SWE, their open-source coding agent, now sends each thread to the cheapest model likely to finish the job.

they tested it on 973 threads, half routed and half always on GPT-6 Astra.

median cost per thread: $0.94 against $2.61. that's 64% lower.

threads ending in a merged PR: 29.2% routed, 27.3% on Astra. LangChain's p-value for that gap is 0.49.

the router sent 56% of threads to the balanced model, 34% to the fast one and 10% to the performance one.

they also tried always using the fast model. that test ended within a day, because engineers reported low output quality.

LangChain's own caveats: the control was the most expensive model, and the router is a proof of concept.

Which three models would you put in the tiers?

🔗 View Quoted Tweet

💬72🔄17❤️152👀11121📊74

⚡ Powered by xgo.ing

we are thinking a lot about how to make deepagents more async native. specifically about making asyn...

Harrison Chase (@hwchase17) — Tue, 13 Oct 2026 09:23:50 GMT

we are thinking a lot about how to make deepagents more async native. specifically about making async subagents work better/more prominent

one key question is how does a finished subagent "notify" the parent agent? some default behaviors we're considering for when the parent is still running:

1/ interrupt parent run, apply subagent update, continue (basically same as "steer")

2/ wait for parent run to finish, then batch subagent update with any other new messages (basically same as "queue")

another question is how to handle human in the loop? what if main agent is not running but paused and waiting explicit human response to an action?

a/ wait for human to respond, then batch the subagent update with that

b/ resume thread (which would cancel the human interrupt) with the subagent update (thread could then re-ask for the same human approval)

any other options we're missing?

💬135🔄5❤️134👀20494📊133

⚡ Powered by xgo.ing

Ha gave this take to our team last week Will share some (very basic) stuff I vibe coded around this...

Harrison Chase (@hwchase17) — Mon, 12 Oct 2026 10:03:36 GMT

Ha gave this take to our team last week

Will share some (very basic) stuff I vibe coded around this shortly

Lots to be explored here!

Nicolas Bustamante@nicbstme

One of my more controversial bets: the AI feed will be the form factor of 2027!

Most people don’t know what to prompt, and they won’t prompt their way into complicated agentic workflows. With chat, we ask users for digital labor: figure out what AI can do, explain the task, provide the context, repeat. That’s a lot to ask if we want billions of people using agents every day. 

Meanwhile, people already spend hours in feeds. Twitter, Instagram, LinkedIn. They know how to scroll, discover something useful, and act on it.

An intelligent system that knows you should bring you the next useful action. Here’s the email you need to answer, already drafted. Here are the latest sales numbers, with an update ready to send to your team. You approve, edit, dismiss, or jump into chat to go deeper. That’s how people will discover what agents can do, without having to invent a prompt first.

And you can put ads in a feed, one of the biggest business models on the internet. A system that understands your plans can recommend a relevant restaurant, product, or service, clearly labeled as sponsored. You get a reason to come back every day, useful work ready to approve, and a business model to fund the agents doing it.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬60🔄10❤️254👀41646📊105

⚡ Powered by xgo.ing

this is the summary slide that describes how all agents in the future will look very much aligns wi...

Harrison Chase (@hwchase17) — Sun, 11 Oct 2026 15:38:40 GMT

this is the summary slide that describes how all agents in the future will look

very much aligns with our vision for managed deepagents docs.langchain.com/langsmith/pyth…

durable workflows: builds on LangGraph/LSD docs.langchain.com/langsmith/depl…

can use a computer: uses Langsmith Sandboxes docs.langchain.com/langsmith/pyth…

memory and skills: langsmith context hub docs.langchain.com/langsmith/pyth…

Triggers and scheduled jobs: schedules docs.langchain.com/langsmith/pyth…

connectors and auth docs.langchain.com/langsmith/pyth… 

text and voice channels docs.langchain.com/langsmith/pyth…

payments and identity - only one missing!

evals docs.langchain.com/langsmith/pyth…

observability: massively, through langsmith docs.langchain.com/langsmith/obse…

Lee Robinson@leerob

How do always-on, proactive agents like @Bot work?

Watch my lecture at Stanford CS146S:

35 How we got here
13 What changed in the models
06 Inside an always-on agent
07 The harness
26 Context engineering
26 Where this is going

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬57🔄40❤️392👀23184📊136

⚡ Powered by xgo.ing

this is a great talk if you want to just follow along in slides, extracted screenshots here to make...

Harrison Chase (@hwchase17) — Sun, 11 Oct 2026 14:57:04 GMT

this is a great talk

if you want to just follow along in slides, extracted screenshots here to make it easy

writing up thoughts as we speak

…p6ke32btx7kfwxvhewash4vi.smithbox.dev/ZXlKaGJHY2lPaU…

Lee Robinson@leerob

How do always-on, proactive agents like @Bot work?

Watch my lecture at Stanford CS146S:

35 How we got here
13 What changed in the models
06 Inside an always-on agent
07 The harness
26 Context engineering
26 Where this is going

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬18🔄27❤️266👀30402📊93

⚡ Powered by xgo.ing

historically, most of our internal agents (openswe for coding, gtm agent for outreach) have acted "o...

Harrison Chase (@hwchase17) — Sun, 11 Oct 2026 13:09:43 GMT

historically, most of our internal agents (openswe for coding, gtm agent for outreach) have acted "on behalf of" whoever is using them

but as they start to do more proactive, background work... I wonder if that should change. eg i really like the idea Hone put forward of agents that are striving towards a goal (x.com/hone/status/21…)... these kind of seem like coworkers, and i wonder if they should be treated as such

Harrison Chase@hwchase17

what is the future of building internal company agents? how will their identity work?

fixed set of credentials
"on behalf of" flow (uses credentials of whoever is using it)
some fixed credentials, others "on behalf of"

🔗 View Quoted Tweet

💬51🔄1❤️108👀19228📊49

⚡ Powered by xgo.ing

what is the future of building internal company agents? how will their identity work? 1. fixed set ...

Harrison Chase (@hwchase17) — Sun, 11 Oct 2026 13:05:59 GMT

what is the future of building internal company agents? how will their identity work?

fixed set of credentials
"on behalf of" flow (uses credentials of whoever is using it)
some fixed credentials, others "on behalf of"

💬99🔄5❤️62👀21607📊92

⚡ Powered by xgo.ing

agree - most orchestration steps don't need a frontier model for Open SWE we moved model choice int...

Harrison Chase (@hwchase17) — Sat, 10 Oct 2026 04:31:37 GMT

agree - most orchestration steps don't need a frontier model

for Open SWE we moved model choice into the harness, so each task goes to the cheapest model that still does the job, tested against quality. median cost per task dropped 64%

langchain.com/blog/how-to-bu…x.com/yuhasbeentaken…

Yum⋆₊˚@yuhasbeentaken

the more i use deepseek v4.1 flash, the harder it is to justify using a frontier model for everything.

a few more things i’ve noticed:

1. it’s insanely good as an orchestrator. i can let it coordinate coding tasks, testing, research, and subagents without burning frontier-model money.

2. once the plan is clear, it handles repetitive implementation work extremely well.

3. i’ve started using it for verification too: reviewing code changes, running checks, and catching obvious mistakes is cheap enough to do constantly.

i still bring in opus or sol for harder edge cases...

but deepseek is doing more and more of the actual work.

🔗 View Quoted Tweet

💬77🔄10❤️107👀12338📊67

⚡ Powered by xgo.ing

good framing from @ch3nweiii - a lot of agent steps are yes/no calls, not generation Jev answers th...

Harrison Chase (@hwchase17) — Sat, 10 Oct 2026 04:31:10 GMT

good framing from @ch3nweiii - a lot of agent steps are yes/no calls, not generation

Jev answers those as typed questions with calibrated probabilities, so the frontier model sticks to the hard work

where those calls sit in a harness: langchain.com/blog/building-…x.com/ch3nweiii/stat…

Chen@ch3nweiii

whoever built this realized we've been using our smartest AI for the dumbest possible jobs

your $20-$200/mo frontier model is researching, writing code, planning…

and then you're paying that same brain to decide YES / NO.

Jev flips that.

it's a tiny decision model that sits in front of the expensive stuff and takes the boring forks:

-> which agent goes next?
-> is this worth researching?
-> reply or ignore?
-> publish, wait or escalate?
-> BUY / SELL / HOLD?
-> click this or keep looking?

then I saw the numbers people are posting and the whole architecture started making a lot more sense.

~1,000 papers classified for around $0.08
~500 emails sorted for around $0.035

browser-agent loops shown at around 7 seconds for ~$0.0039

and at the quoted pricing, roughly 10,000 ~1K-token decisions comes out around $0.42.

that's not because Jev got smarter than Claude.

it's because nobody asked it to be Claude.

the setup is stupidly simple:

-> big LLM gets the hard thinking, research and writing
-> Jev gets the thousands of tiny decisions between those steps
-> code actually clicks, saves, sends and runs

and suddenly you start seeing these little decisions everywhere.

> your inbox.
> X feed.
> support queue.
> meetings.
> leads.
> video clips.
> browser agents.
> even paper-trading loops.

that's the part I think most agent builders are missing.

the expensive model doesn't need to touch every step just because it's able to touch every step.

> let Claude think.
> let Jev decide what deserves Claude.
> let code do the actual work.

I broke down how to wire Jev into your agent + the exact router setup in the article below

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬32🔄16❤️177👀20906📊65

⚡ Powered by xgo.ing

This is great point Goodharts law: when a measure becomes a target, it ceases to be a good measure ...

Harrison Chase (@hwchase17) — Fri, 09 Oct 2026 08:28:18 GMT

This is great point

Goodharts law: when a measure becomes a target, it ceases to be a good measure

Initial take: eval driven development works for narrowly scoped things, but for more autonomous agents it doesn’t

So then - how do you hill climb those?

Hunter Gerlach@HunterGerlach

Create new high-value evals
Realize you need to fight against Goodhart's Law
Go to 1

🔗 View Quoted Tweet

💬77🔄12❤️130👀11182📊72

⚡ Powered by xgo.ing

eval driven development

Harrison Chase (@hwchase17) — Fri, 09 Oct 2026 07:50:53 GMT

eval driven development

Jerry Liu@jerryjliu0

These days, you can pretty much solve any task by defining an eval and hillclimbing over it, instead of directly defining the deterministic/agentic workflow to solve it.

The data provider companies' entire job is to define evals for all economic activity to make the frontier models capable of doing anything. 

Your job then becomes pointing frontier intelligence in the right direction - defining what to solve, and how to measure what good looks like. 

Agent application interfaces will evolve to capture this. The most complex processes will still need some sort of explicit workflow builder interface, but most tasks can be compressed into goals and eval instructions.

🔗 View Quoted Tweet

💬19🔄21❤️269👀29518📊59

⚡ Powered by xgo.ing

trajectory labeling is really several questions, not one pass/fail - nice framing here that's how J...

Harrison Chase (@hwchase17) — Fri, 09 Oct 2026 07:06:01 GMT

trajectory labeling is really several questions, not one pass/fail - nice framing here

that's how Jev as a judge works in LangSmith evals: difficulty and correctness each get their own typed answer, in one pass, on every trace

langchain.com/blog/jev-is-no…x.com/xennygrimmato_…

Vaibhav Tulsyan@xennygrimmato_

This is a very clever usecase for Jev-like decision models. Trajectory labelling at scale is key for RSI.

Why is that?
Every trajectory needs to be assessed across several dimensions.
1. How difficult was the task?
2. Was the agent’s solution correct?
3. Did the agent consider a diverse set of design choices before deciding to pick a specific route?
4. Is the agent’s thinking directionally correct?

Imagine answering these questions for millions of agents each of which spawns tens or hundreds of subagents! Token cost for trajectory labelling needs to be driven down a lot to do this at scale.

This work is unique and solves a rather under-looked aspect of RL research.

🔗 View Quoted Tweet

💬39🔄6❤️116👀11298📊47

⚡ Powered by xgo.ing

Sam on where decision models fit in a harness and how to use Jev with LangChain https://t.co/TgYoOi...

Harrison Chase (@hwchase17) — Thu, 08 Oct 2026 19:03:23 GMT

Sam on where decision models fit in a harness and how to use Jev with LangChain

langchain.com/blog/building-…x.com/jaminball/stat…

Jamin Ball@jaminball

Next up on First Pass: a conversation with
@samecrowder from @LangChain
on decision models.   

Jev from @typesafeai  took the world by storm. Then OpenAI and Databricks released producs (Decisions API and ai_decide function)  So what are Decision Models? We dig in!

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬20🔄9❤️32👀6643📊19

⚡ Powered by xgo.ing

interrupt london is next tuesday really looking forward to hearing from Orange and OVO on what's ac...

Harrison Chase (@hwchase17) — Wed, 07 Oct 2026 19:51:15 GMT

interrupt london is next tuesday

really looking forward to hearing from Orange and OVO on what's actually working with agents in production. we're also running hands-on workshops on Deep Agents and Engine

x.com/LangChain/stat…

LangChain@LangChain

Only one week to go until Interrupt London!

interrupt.langchain.com/london

🔗 View Quoted Tweet

💬15🔄8❤️43👀6160📊14

⚡ Powered by xgo.ing

One cool thing deepagents now supports: dynamically loading tools when is a skill is loaded Some sk...

Harrison Chase (@hwchase17) — Wed, 07 Oct 2026 18:58:29 GMT

One cool thing deepagents now supports: dynamically loading tools when is a skill is loaded

Some skills require specific tools that you may not always want to have available. This now allows you to pair them 

With openai/anthropic models, it’s possible to do this in a way where it doesn’t break the prompt cache!

Sydney Runkle@sydneyrunkle

x.com/i/article/2107…

🔗 View Quoted Tweet

💬35🔄12❤️87👀9475📊37

⚡ Powered by xgo.ing

We have a really strong team :) Come join us!

Harrison Chase (@hwchase17) — Wed, 07 Oct 2026 09:34:55 GMT

We have a really strong team :)

Come join us!

💬7🔄6❤️91👀17761📊20

⚡ Powered by xgo.ing

Good take on harnesses

Harrison Chase (@hwchase17) — Tue, 06 Oct 2026 04:26:37 GMT

Good take on harnesses

David Cramer@zeeg

You’re living in a bubble. You need to understand that a whole world lives beyond the constraints you have manufactured or imagined.

For example the codex harness provides very little in terms of actual functionality - the same goes for most coding harnesses. They will be generally superseded by more capable specialized harnesses. Even if you scoped codex to only a coding-focused harness - which it really isn’t - it is lacking compared to purpose built coding harnesses that compete with it. 

Right now labs have the advantage and can somewhat dictate terms but you should not believe that will be true forever. There’s seemingly a belief that we should be thankful for the (limited) capabilities we’ve been given- that we should allow a few companies in the world to dictate what we can and can’t do with technology. That’s never worked before and nothing makes this tech any more defensible. 

In five years we will almost certainly be consuming local models for a majority of daily tasks, and for the things that can’t be done locally it’s going to be entirely commoditized. Once that’s true it means vendors will have to start competing by giving up more to the customer rather than the increasing restrictions being done today.

🔗 View Quoted Tweet

💬22🔄1❤️83👀20251📊32

⚡ Powered by xgo.ing

cool to see cognition doing "dreaming" - memory needs an offline cleanup loop, not just better retri...

Harrison Chase (@hwchase17) — Mon, 05 Oct 2026 20:28:50 GMT

cool to see cognition doing "dreaming" - memory needs an offline cleanup loop, not just better retrieval, otherwise stale records pile up

open q is how inferred memories get validated before the agent uses them

glad it's an open standard too - memory should be open

x.com/cognition/stat…

Cognition@cognition

Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work

At night Devin self-improves its memory to remove stale records and discover latent information

We are creating an OSS standard called Agent Memory Repo: cognition.com/agent-memory-r…

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬69🔄9❤️148👀16152📊72

⚡ Powered by xgo.ing

we're seeing our coding agent costs decrease significantly for second month in row longer blog on t...

Harrison Chase (@hwchase17) — Sun, 04 Oct 2026 10:39:18 GMT

we're seeing our coding agent costs decrease significantly for second month in row

longer blog on this later, but three quick steps we took

1/ cost visibility. you cant control what you cant measure. Start tracking everything to langsmith. This gives us super granular visibility into who is using what, how they are using it, etc. LangSmith supports first party integrations with all main coding harnesses langchain.com/blog/your-codi…

2/ cost controls. Through our LLM gateway, set user level cost caps. (You can increase your cap by talking to our VPE). This helps curb unexpected or accidental spend. docs.langchain.com/langsmith/llm-…

3/ optimize the harness. Obviously 3rd party harnesses are tougher to optimize, but we are doing more and more coding through OpenSWE, our open source cloud agent harness. Things like model routing help bring the cost down langchain.com/blog/how-to-bu…

feel free to reach out if you want help implementing these at your co!

💬97🔄15❤️180👀18685📊117

⚡ Powered by xgo.ing

specialized agents or one mega agent? we're having this debate internally...

Harrison Chase (@hwchase17) — Sat, 03 Oct 2026 20:05:50 GMT

specialized agents or one mega agent?

we're having this debate internally...

a16z@a16z

.@OpenRouter co-founder Alex Atallah on personal agents: 10 specialized chiefs of staff beat one universal agent. One job each means you can see exactly where it's failing.

"The worst part about doing cross-domain joins with your personal agent is that the more work you give it, the more understanding of what's going on you're sacrificing."

"And no one is taking responsibility for that sacrifice. Agents don't have any responsibility."

"With a universal agent that's doing all things, I can't adjust how much understanding I'm sacrificing."

"It points me towards: maybe down the road, the sub-agents people use will be very vertically focused, and we have a chief of staff type agent that coordinates between them."

"You need vertically focused agents where this agent is more responsible, psychologically, for these things. I want quality checks that make sure it's doing those things correctly. It doesn't need to focus on anything else, it has one focus area."

"I wonder if that's going to help people get at least a loose sense of responsibility on top of agents."

@alexatallah

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬95🔄11❤️189👀27720📊95

⚡ Powered by xgo.ing

cool project by @dbreunig in langchain this is ModelRouterMiddleware - jev reads the first message,...

Harrison Chase (@hwchase17) — Sat, 03 Oct 2026 20:05:10 GMT

cool project by @dbreunig

in langchain this is ModelRouterMiddleware - jev reads the first message, picks the model, and that model handles the whole run

jev is cheap enough that you can also re-pick after each tool result with a custom hook

docs: docs.langchain.com/oss/python/int…x.com/dbreunig/statu…

Drew Breunig@dbreunig

A quick example of a model router, using DSPy and Jev, based on your initial prompt. 

Sketched this out in ~15 min, by request, and am surprised at how well it works! gist.github.com/dbreunig/949b4…

🔗 View Quoted Tweet

💬21🔄15❤️89👀10906📊31

⚡ Powered by xgo.ing

good pattern from this google paper the verifier starts from a clean context, so the explorer's rea...

Harrison Chase (@hwchase17) — Fri, 02 Oct 2026 17:42:10 GMT

good pattern from this google paper

the verifier starts from a clean context, so the explorer's reasoning can't talk it into a bad proof

so the explorer can try wild ideas and the verifier tosses the wrong ones

same reason deepagents subagents get their own context

x.com/omarsar0/statu…

elvis@omarsar0

Banger paper from Google Research on multi-agent proof discovery.

(bookmark it)

It's really interesting to see this emerging multi-agent pattern: not enforcing too much execution structure and pairing it with dedicated agents for advising and verification.

I think it is generally applicable as well. Great read.

Here is how it works:

Cogentic runs on Gemini and works on open problems in theoretical computer science, starting from the problem statement with no expert hints.

The system works in rounds, and the orchestrator decides how many provers to run in each round. Every prover gets one direction to work on, such as a specific bound or a counterexample search, plus a short briefing that a summarizer agent writes from earlier attempts and verifier feedback. Each summarizer writes its briefing independently, so provers in the same round read different summaries of the same history.

Each draft goes through two adversarial verifiers. One checks the draft on its own, and the other reads all of the round's drafts side by side to catch shared mistakes. A draft is accepted only if both pass it.

The agents share state through two disk documents. A record logs every attempt with the objection it failed on, and a ledger stores verified lemmas and ruled-out directions. An auditor extracts correct lemmas from rejected proofs, verifies them again independently, and adds them to the ledger.

A separate process advisor reads the verification logs across rounds and updates the instructions given to provers and verifiers. The orchestrator and the advisor can't give mathematical opinions, so the provers provide all the math.

It produced new results on five open problems in online learning, auction theory, and mechanism design, each checked by domain experts. Most problems took around 100 Gemini calls, and the hardest took around 1,000.

Paper: arxiv.org/abs/2609.40324

Chat with Paper: academy.dair.ai/papers/cogenti…

🔗 View Quoted Tweet

💬35🔄9❤️99👀9839📊42

⚡ Powered by xgo.ing

every agent team ends up wanting its own review ui over their traces - LangSmith Custom Apps lets yo...

Harrison Chase (@hwchase17) — Fri, 02 Oct 2026 17:33:45 GMT

every agent team ends up wanting its own review ui over their traces - LangSmith Custom Apps lets you build it on your own data and ship it right into the workspace

x.com/LangChain/stat…

LangChain@LangChain

💻 Build the interface you want w/ your LangSmith data
✅ Publish it into your workspace
🙅 Skip hosting, auth, + permissions work

Get started with Custom Apps today.langchain.com/blog/langsmith…nW

🔗 View Quoted Tweet

💬16🔄2❤️46👀6676📊16

⚡ Powered by xgo.ing

Memory is hard to get working well, especially in a company setting Making some improvements to how...

Harrison Chase (@hwchase17) — Fri, 02 Oct 2026 16:13:18 GMT

Memory is hard to get working well, especially in a company setting

Making some improvements to how we memory in managed deepagents!

Caspar@caspar_br

User memory in Managed Deep Agents 0.8: your agent can remember the people it works with langch.in/mda

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬27🔄9❤️93👀12466📊34

⚡ Powered by xgo.ing

congrats @AravSrinivas! more entrants in the decision model class, and this one is open weights che...

Harrison Chase (@hwchase17) — Thu, 01 Oct 2026 23:45:36 GMT

congrats @AravSrinivas! more entrants in the decision model class, and this one is open weights

cheap typed answers for the small calls inside a harness (routing, approvals, judging), big model for the rest

x.com/AravSrinivas/s…

Aravind Srinivas@AravSrinivas

We’re open sourcing a state of the art multimodal Decision Model, pplx-decider-27b, and are offering it in a new Decisions API at 4 cents per million input tokens and free output tokens. We intend to bring down the price even further over the coming days. Enjoy!

🔗 View Quoted Tweet

💬18🔄8❤️64👀10533📊22

⚡ Powered by xgo.ing

pretty clear every agent harness needs a durable runtime pi :: pi-durable deepagents :: langgraph

Harrison Chase (@hwchase17) — Thu, 01 Oct 2026 22:45:58 GMT

pretty clear every agent harness needs a durable runtime

pi :: pi-durable

deepagents :: langgraph

Pi@pidotdev

People of Pi: We've shipped Pi 1.0 with Pi Durable. Go make them yours.

earendil.com/posts/pi-1-0/

🔗 View Quoted Tweet

💬44🔄15❤️303👀38680📊80

⚡ Powered by xgo.ing

decision model season indeed said there'd be more entrants before long - now open weights too, nice...

Harrison Chase (@hwchase17) — Thu, 01 Oct 2026 17:37:40 GMT

decision model season indeed

said there'd be more entrants before long - now open weights too, nice work @ritakozlov + the cloudflare team

your harness should be able to swap its decision model as easily as its main model

 x.com/ritakozlov/sta…

rita kozlov 🐀@ritakozlov

it's decision model season 🍂

today, we open sourced clef and clef-flash, our first homegrown models — both now available on @cloudflare workers ai! 

demo / leaderboard:clef-evals.workers-ai-mle.workers.devk

blog:blog.cloudflare.com/clef-decision-…z

🔗 View Quoted Tweet

💬20🔄11❤️114👀14423📊32

⚡ Powered by xgo.ing

model routing - "no one knows what it means, but its provocative, gets the people going" we try to ...

Harrison Chase (@hwchase17) — Thu, 01 Oct 2026 17:24:18 GMT

model routing - "no one knows what it means, but its provocative, gets the people going"

we try to shed light on how to do it properly

1. understand tasks
2. understand the models
3. build the router **in the harness**
4. track outcomes

lower costs, with no performance hit

Sydney Runkle@sydneyrunkle

x.com/i/article/2105…

🔗 View Quoted Tweet

💬33🔄17❤️149👀20266📊54

⚡ Powered by xgo.ing

great writeup by @pirroh harness doesn't disappear as models get smarter - it shifts from deciding ...

Harrison Chase (@hwchase17) — Tue, 29 Sep 2026 21:25:27 GMT

great writeup by @pirroh

harness doesn't disappear as models get smarter - it shifts from deciding to offering

in deepagents you give each subagent its own model, tools, and permissions - the main loop picks which to spawn and when

docs: docs.langchain.com/oss/python/dee…

Michele Catasta@pirroh

x.com/i/article/2104…

🔗 View Quoted Tweet

💬46🔄21❤️137👀19777📊61

⚡ Powered by xgo.ing

listen is a great company, alfred and florian are awesome founders the problem space listen is tack...

Harrison Chase (@hwchase17) — Tue, 29 Sep 2026 21:24:48 GMT

listen is a great company, alfred and florian are awesome founders

the problem space listen is tackling is also an amazing use case for AI, very bullish on what they will still build here!

Alfred Wahlforss@itsalfredw

Listen Labs is joining @Salesforce!

A year and a half ago we launched Listen, an AI platform for understanding customers, and it’s been a crazy ride ever since.

Today, we work with some of the largest companies in the world, including Microsoft, Anthropic, and Sweetgreen.

What started as an AI interviewer is now a full platform for understanding customers. Listen finds the right people, interviews them, analyzes what they say, and even simulates how they'll behave.

Listen’s growth quickly accelerated and while we were raising our next round, we met @Benioff.

Marc is a hero of mine. I've taken countless ideas from Behind the Cloud and watched him framemog every AI CEO on the planet. We're honored to get framemogged next.

With Salesforce, we can bring Listen to every company in the world, much faster.

To our customers, our mission remains the same.I stay CEO, @Florian stays CTO and the same team will keep building the product you rely on, now inside Salesforce AI Labs. Thank you for believing in us early.

To our team, families, and everyone who believed in us before there was much to believe in, thank you.

These years have been the most fun years of my life, and it’s still only just the beginning for Listen.

Now back to work.  🚀

🔗 View Quoted Tweet

💬10🔄1❤️47👀8420📊11

⚡ Powered by xgo.ing

No

Harrison Chase (@hwchase17) — Tue, 29 Sep 2026 18:19:39 GMT

No

Rhys@RhysSullivan

it's crazy how far out the labs are building at the application layer

on the one hand it makes sense, on the other hand do you really want all of your company knowledge, docs, workflows built so locked in to a single model?

🔗 View Quoted Tweet

💬9🔄3❤️73👀19867📊13

⚡ Powered by xgo.ing

company OS (org harness) vs personal agent, similarities and differences: differences: - Many princ...

Harrison Chase (@hwchase17) — Tue, 29 Sep 2026 14:25:29 GMT

company OS (org harness) vs personal agent, similarities and differences:

differences:
- Many principals (users), one actor (agent): will need to be multiplayer, handle multiple users (sometimes even in same thread), and as a result handle auth/memory correctly
- governance: observability, auditability, admin control plane matter way more
- interaction pattern: more event driven and async

similarities:
- should be able to write/execute code
- browser use likely very important! (key difference from a coding harness)
- skills/MCP as standards
- core agent loop/logic

what am i missing?

💬76🔄12❤️200👀14440📊111

⚡ Powered by xgo.ing

coding harness -> personal harness -> org harness (company harness? enterprise harness? still...

Harrison Chase (@hwchase17) — Mon, 28 Sep 2026 22:36:10 GMT

coding harness -> personal harness -> org harness

(company harness? enterprise harness? still figuring out right name, open to suggestions)

💬70🔄1❤️176👀18079📊87

⚡ Powered by xgo.ing

every agent will need a sandbox and the harness should live outside it - separate the brains from t...

Harrison Chase (@hwchase17) — Mon, 28 Sep 2026 22:13:33 GMT

every agent will need a sandbox

and the harness should live outside it - separate the brains from the hands

great to see nvidia open sourcing tools here, agree with Andrew this is part of the harness not a bolt on

x.com/AndrewYNg/stat…

Andrew Ng@AndrewYNg

The OpenAI-Hugging Face hack was enabled by weak sandboxing. It is great that Nvidia is releasing open source tools for sandboxing AI agents. OpenWorker, our open-source agent harness supporting cybersecurity workflows, is proud to support this.

A sandbox gives an agent limited permissions. OpenWorker is building on Nvidia OpenShell and will support running each agent's commands inside a sandbox. Only the files relevant to the task go in. Secret API keys, your web browser login credentials, the ability to access arbitrary websites, are inaccessible to the agent by default. These restrictions are implemented in deterministic code rather than by prompting an LLM, which can make mistakes or be susceptible to prompt injections. Further, all actions are logged for monitoring and audit.

I'm grateful for @JensenHuang's leadership making AI agents more secure. OpenWorker (which @rohitcprasad and I are working on) will continue to improve security for agents.

🔗 View Quoted Tweet

💬44🔄12❤️136👀15855📊52

⚡ Powered by xgo.ing

retrieval is a decision problem, not just a similarity problem cool experiment from @assaf_elovic s...

Harrison Chase (@hwchase17) — Mon, 28 Sep 2026 16:32:40 GMT

retrieval is a decision problem, not just a similarity problem

cool experiment from @assaf_elovic swapping embeddings for Jev in GPT Researcher - 73% vs 46% relevant context

decision models will show up all over the harness

x.com/assaf_elovic/s…

Assaf Elovic@assaf_elovic

Didn't expect this 🤯

We replaced embeddings with Jev in GPT Researcher's RAG pipeline and tested both on 28 research tasks from SimpleQA and open ended research.

Jev beat embeddings on every quality measure we ran:
- 59% more relevant context (73% vs 46%)
- Reports preferred 15 to 3 in blind comparisons
- Same cost per report

GPT Researcher now runs on Jev by default, and no longer needs embeddings at all.

All you need is @LangChain + @tavilyai +Jev for the perfect RAG system.

Check out the repo here:github.com/assafelovic/gp…q
Research:docs.gptr.dev/docs/gpt-resea…q

🔗 View Quoted Tweet

💬48🔄41❤️317👀46506📊100

⚡ Powered by xgo.ing

> feels way more shippable than "the LLM decided" probabilities are the underrated part of decis...

Harrison Chase (@hwchase17) — Sun, 27 Sep 2026 04:39:20 GMT

> feels way more shippable than "the LLM decided"

probabilities are the underrated part of decision models

you can threshold them, log them, write evals against them

 https://t.co/k4eRrMgNV0

Hiếu Nguyễn Trung@Hiumonday

@hwchase17 honestly the probabilities are the best part. you can set a threshold, log it, write tests for it. feels way more shippable than "the LLM decided"

🔗 View Quoted Tweet

💬0🔄0❤️0👀140

⚡ Powered by xgo.ing

jev and langgraph are a great combo! modeling agents as complex systems, with AI imbued in them, ma...

Harrison Chase (@hwchase17) — Fri, 25 Sep 2026 19:07:11 GMT

jev and langgraph are a great combo!

modeling agents as complex systems, with AI imbued in them, makes a ton of sense

langgraph is the best way to model them as such, and using Jev inside langgraph to turbo charge all the small decisions that need to be made is awesome combo

Sydney Runkle@sydneyrunkle

x.com/i/article/2103…

🔗 View Quoted Tweet

💬47🔄32❤️230👀26118📊72

⚡ Powered by xgo.ing

Proactive agents run in the background, schedules are (one way) how you achieve that!

Harrison Chase (@hwchase17) — Wed, 23 Sep 2026 16:47:23 GMT

Proactive agents run in the background, schedules are (one way) how you achieve that!

LangChain@LangChain

You can add schedules to Managed Deep Agents.

1️⃣ Add a schedule file to schedules/ with a cron expression + a prompt
2️⃣ Deploy
...And the agent runs itself

💻 Docs:docs.langchain.com/langsmith/pyth…M

⏯️ Demo from @ndrezn ⤵️

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬48🔄5❤️62👀11311📊40

⚡ Powered by xgo.ing

🖼️ Improved UI for Jev models in LangSmith agents are going to complex systems of many llms calls ...

Harrison Chase (@hwchase17) — Tue, 22 Sep 2026 18:49:55 GMT

🖼️ Improved UI for Jev models in LangSmith

agents are going to complex systems of many llms calls

many llm calls will be jev-like (decision models)

we've updated tracing view in LangSmith to display these types of calls (state, questions, choices, output) more clearly!

💬44🔄10❤️91👀22056📊51

⚡ Powered by xgo.ing

this is cool kind of like "goal" but with multiple in a row you need to hit?

Harrison Chase (@hwchase17) — Mon, 21 Sep 2026 20:47:09 GMT

this is cool

kind of like "goal" but with multiple in a row you need to hit?

Josh Rosen@JoshARosen

x.com/i/article/2102…

🔗 View Quoted Tweet

💬20🔄15❤️119👀20657📊41

⚡ Powered by xgo.ing

try out jev in langsmith today! can use it to score traces cheap and accurately

Harrison Chase (@hwchase17) — Mon, 21 Sep 2026 17:30:00 GMT

try out jev in langsmith today! can use it to score traces cheap and accurately

💬13🔄7❤️36👀13307📊17

⚡ Powered by xgo.ing

the reason i'm excited about decision models: agentic systems are just good engineering around mode...

Harrison Chase (@hwchase17) — Mon, 21 Sep 2026 17:17:01 GMT

the reason i'm excited about decision models:

agentic systems are just good engineering around models. "build prod, not god"

these types of fast, cheap & accurate models are super important for that

the excitement around jev shows that most people think this way as well!

💬39🔄6❤️85👀13331📊46

⚡ Powered by xgo.ing

semif is pretty close to jev (75.4 to 74.7) on JevBench from @airesearch12 lots of other open sour...

Harrison Chase (@hwchase17) — Mon, 21 Sep 2026 16:50:24 GMT

semif is pretty close to jev (75.4 to 74.7) on JevBench from @airesearch12 

lots of other open source options as well!

x.com/airesearch12/s…

Harrison Chase@hwchase17

jev is a new type of model - "decision models"

there are a few open source alternatives, one is SemIf (based on qwen3.5)

we're hosting SemIf and offering it for free through LangSmith Gateway for the next week!

🔗 View Quoted Tweet

💬13🔄9❤️86👀10715📊22

⚡ Powered by xgo.ing

jev is a new type of model - "decision models" there are a few open source alternatives, one is Sem...

Harrison Chase (@hwchase17) — Mon, 21 Sep 2026 15:59:16 GMT

jev is a new type of model - "decision models"

there are a few open source alternatives, one is SemIf (based on qwen3.5)

we're hosting SemIf and offering it for free through LangSmith Gateway for the next week!

Jacob Lee@Hacubu

🆓 Free Open-Source Jev 🤝 LangSmith Gateway

Decision models are becoming a first-class part of the LangSmith Gateway!

We're serving SemIf, an open-source decision model, free for the next week.

Compatible with the TypeSafe SDK - just change a string to try it!

Docs in 🧵

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬27🔄10❤️140👀35682📊37

⚡ Powered by xgo.ing

Decision Model is a great name, and I expect other entrants in this new model class before long

Harrison Chase (@hwchase17) — Sun, 20 Sep 2026 16:31:25 GMT

Decision Model is a great name, and I expect other entrants in this new model class before long

Maggie Appleton@Mappletons

Can we agree to call Jev a "Decision Model"?
I feel like they buried this in the explainations and docs.
Language models output language. Decision models output decisions.

🔗 View Quoted Tweet

💬38🔄5❤️102👀16799📊36

⚡ Powered by xgo.ing

"jev as a judge" cheap and fast semantic verifiers! great for evals - especially online evals where...

Harrison Chase (@hwchase17) — Sun, 20 Sep 2026 13:53:23 GMT

"jev as a judge"

cheap and fast semantic verifiers! great for evals - especially online evals where you want to grade LOTS of traces

LangChain@LangChain

We tested Jev against LLM judges on accuracy, repeatability, latency, and cost to see whether System One models could offer a new approach to agent evaluation. x.com/i/article/2101…

🔗 View Quoted Tweet

💬56🔄46❤️498👀69460📊129