Harrison Chase

If you like evals, this is a good role for you

RT Brace i need to hire someone who's really really good at evals and benchmarks in SF (or maybe nyc if you're super special) if this is you, or you know someone who fits this, please dm me! i'd love to chat

RT Paul Iusztin "Memory for agents is still early, with little to no standards... but one common pattern is emerging: wiki memory." This is how @hwchase17 (co-founder of @LangChain) recently described the current state of agent memory. Over the past year, I've been building two separate systems that accidentally converged on the same idea. One builds project-scoped LLM wikis from my notes, research, and conversations. The other serves my unified memory through a @fastmcp server. But I think combining them is the missing piece. Today, most AI agents repeatedly search the same documents, retrieve the same chunks, and reconstruct the same understanding every time you ask a question. Wiki memory changes that abstraction. Instead of reasoning directly over raw documents, the agent incrementally builds a small, structured Markdown wiki from your unified memory that becomes its local working knowledge. The first conversation creates it. Every conversation after that improves it. This is a different model from traditional RAG. Today, I only use LLM wikis for personal projects built on top of my Second Brain. Now imagine doing the same thing for enterprise data… Your unified memory remains the source of truth. The wiki becomes a lightweight, task-specific layer the agent can reason over. @fastmcp (by @PrefectIO) sits between the two. It handles the business logic: Search the unified memory Build the wiki Synchronize changed sources Persist new knowledge The harness simply calls those tools and skills. Even better, this doesn't require GraphRAG across your entire corpus. So you: Keep your existing data sources as they are (with or without KGs) Build a local wiki that implements a light KG via file references Point an agent directly to the wiki which uses progressive disclosure to query it The MCP server acts only as the engine between your company data and the LLM wiki. The harder engineering problem is keeping it synchronized. Every page already know...

RT Sidra Miconi . @hwchase17 OpenWiki 0.2 is turning documentation from a pile of Markdown into infrastructure that agents can reliably navigate. Adopting the OKF spec adds structured metadata, deterministic indexes, and update logs to every repo wiki. That sounds simple, but it solves a real scaling problem: once documentation reaches hundreds of files, asking an agent to search everything semantically is slow, expensive, and often unnecessary. With tags, categories, descriptions, and canonical resources, agents can retrieve the right context directly. And logs.md means they can inspect what changed instead of rereading the entire knowledge base after every update. The broader idea is useful: agent memory should not depend entirely on fuzzy retrieval. Some knowledge is better expressed as explicit, inspectable structure. Open formats matter here too. If repo knowledge follows a shared specification, viewers, linters, search tools, and coding agents can all operate on the same layer. Better agents will need better models, but they will also need codebases that explain themselves in a machine-readable way. OpenWiki is building that missing interface.

RT Wey Gu 古思为 I really liked OpenWiki and happy to see it evolves so fast to be the best index of not just codebase but more form of knowledge. And today OpenWiki impl the Open Knowledge Format(OKF)🥳 btw, Nowledge Mem also supports OKF, the second day the spec draft was released :D

RT Justin Google with design.md and now OKF!!!

RT Factory Our CTO @EnoReyes sat down with Harrison Chase to talk models, harnesses, and what it actually takes to build a software factory.

there needs to be an OPEN standard for memory OKF (Open Knowledge Format) is one such standard excited to announce that we're now using it in OpenWiki

one of the main questions we got about open wiki was whether we would support OKF if you dont know what OKF is - its a standard for knowledge files check out Brace's video!

fun trip down memory lane

RT Sydney Runkle owning intelligence doesn't mean being able to afford the best model it means owning the agent development lifecycle from model agnostic harness -> observability -> learning from each run

RT Colin Francis OpenWiki 0.2 adopts OKF! this is a big deal for two reasons: 1. OKF adds structured metadata to every doc and builds connections through cross links which means faster results, fewer tokens 2. the growing OKF tools ecosystem lets you visualize your wiki and the connections between docs

RT Brace I just published a short YouTube video on OKF in OpenWiki. It's only 3:30 long, so if you're curious what OKF is, what it'll mean for OpenWiki, and why we decided to adopt, you should give it a watch! https://youtu.be/NxJjMvDN6aE

RT Jacob Lee This year @LangChain 🦜 🔗 has almost tripled in size, from ~100 people to over 300! That's a lot of new faces, and I'm often asked: "What made you join early? What was the company like at seven people?" I'm always honest: I wasn't convinced right away. I mean, three years ago, the website looked like this: http://web.archive.org/web/20230303225335/https://langchain.com Raw HTML and two emojis for a logo. When I told my parents I was joining, they asked, pretty directly, if I'd lost my mind. I met @ankush_gola11 about a month after ChatGPT launched and the open-source started getting real traction. I was completely fascinated by the technology, but my reaction to his pitch was "cool - good luck!" In addition to said website, I had just come off six years founding a company, and didn't want to jump back into an early startup. A few weeks later, a client wanted to build an LLM-powered extraction pipeline - nowhere near as trivial then as it is now. I tried an early version of LangChain.js, and it worked great, so I kept exploring and found the section on agents. The magic there completely hooked me. It truly seemed like something out of science fiction. Thankfully, @hwchase17 and Ankush were still looking to round out their founding team, and I got to see how they operated up close. Unlike most startups, LangChain has never had a name recognition problem, and I knew firsthand how easy it was to start chasing GitHub stars and other vanity metrics. But from the start, the team was serious about seeing through the fluff, listening to the community even when tempers flared, and building what actually solved user problems. They weren't always right - I joke that back then we were a six-month-old startup with a year of tech debt - but I admired how quickly they'd throw out popular beliefs and abstractions the moment evidence showed they were wrong or wouldn't scale. Anyone remember ConversationalRetrievalChain? The website itself was emblematic of this ...

RT Brace http://x.com/i/article/2077795070166913027

this is one thing ive realized - owning that feedback loop (the signals, the output) is really code to owning intelligence

RT Alex Olsen The future is open

RT Andrew Nguonly Owning > renting

what does it really mean to "own your intelligence"? 1. can the system compound *on your terms*? the durable advantage is not the model/system at this point in time - it's the flywheel that compounds over time 2. intelligence is no longer just the model - its the system around the model. if the system is opaque or locked away, you are renting intelligence (not owning it) 3. own the harness. integrate it deeply into your systems. make it specific to the task at hand 4. own the context and memory layer. includes user preferences, org knowledge, workflow patterns. if memory is trapped in vendor product, your learning is trapped to 5. own model optionality. you need freedom to route across models, providers, deployment modes. different tasks require different tradeoffs 6. own the economics. intelligence only matters if you can afford to deploy it broadly. sustainable intelligence at scale 7. own the feedback loop. observability tells you what happened, evals tell you if it worked. without this loop you cannot measure the intelligence, you are guessing s/o @nlarusstone @nfcampos for their thoughts on feedback

dcode with open models! (nemotron 3 ultra)

RT LangChain Brand new Max Agency with @FactoryAI CTO & Co-Founder @EnoReyes!

New Max Agency with @FactoryAI CTO & Co-Founder @EnoReyes. This was a particularly fun one. Since we're both building harnesses, we nerded out pretty deep. Awesome conversation on why the harness matters more than the model running underneath it, how Factory built Missions, + more. ⏯️ YouTube: https://www.youtube.com/watch?v=HbUznYhKFOc 🎧 Apple: https://podcasts.apple.com/us/podcast/the-best-ai-agents-cost-less-than-you-think-eno-reyes-factory/id1891551672?i=1000777066738 🎧 Spotify: https://open.spotify.com/episode/5CkX1dYgF89PGs4n3bJZFM

RT LangChain How @Box uses Deep Agents middleware for Box Agent: 1️⃣Citation generation: Streaming of the answer + citation generation happens in parallel to avoid user interruption 2️⃣Prompt caching: Injects caching on multi-turn conversations, reducing cost + latency 3️⃣Context management: Summarizes conversation history after exceeding 170k tokens, preventing context overflow https://www.langchain.com/blog/building-box-ai-how-an-enterprise-content-platform-went-ai-native-with-deep-agents

Good takes from Caspar on why we’re leaning into slack as an interface for agents

RT Sydney Runkle the agents i use the most are in slack that's where most of my valuable discussions are, and thus the most valuable context for my agents!

RT Caspar Broekhuizen http://x.com/i/article/2077443128122880000

It’s super easy to create custom slack agents now Makes it easy to bring intelligence to where the team already lives

RT LangChain Fleet in Slack just got a big upgrade. Now you can bring any Fleet agent into Slack in one click. Give it a custom identity. Use it in channels and threads. Hand off files, approve its actions, and keep all the context of your work in one place. Read more in our blog: https://www.langchain.com/blog/new-in-langsmith-fleet-bring-agents-into-slack-in-one-click

RT Brace Personal Brain from OpenWiki connects directly to your Twitter account which allows it to ingest posts from your feed (yes this is an api they offer!) and bookmarks. Bookmark ingestion is especially useful since we instruct the agent to place extra emphasis on bookmarks during the memory process i use this all the time to save interesting posts i see, which i want my memory agent to know about try out OpenWiki and build your own personal brain 👇

RT Sydney Runkle we just dropped a comprehensive course on deepagents! learn about agent harnesses and how to build one optimized for your project!

RT Sydney Runkle excited about the continual learning angle you have when you can learn from coding agent traces

Open source coding harness!

RT Sure (e/acc) Re @LangChain really outdid themselves with dcode. It’s currently one of my favourite harnesses for coding. 🙌🏽

we did a bunch of work to standardize how coding agents trace to langsmith

RT LangChain Cursor, Copilot, Pi, and OpenCode tracing: Now in LangSmith. Full session observability, no extra instrumentation. ✅Identify, group, and query any coding-agent trace with the same stable keys, regardless of which agent produced it ✅See the full run tree: turns, model calls, tools, and subagents ✅Token usage and cost per session, out of the box Blog by @harisaiharish https://www.langchain.com/blog/your-coding-agents-are-a-black-box-heres-how-to-crack-them-open

RT Brace Who’s actually using OKF? What’re your reasons behind it/thoughts? We’re going to adopt it in OpenWiki for a couple reasons: - it’s a simple spec. Supporting it doesn’t mean changing much (basically just yaml front matter in each .md file) - the front matter spec should make it easier to build & use deterministic search & filtering tools against the docs (keyword, full text, etc) - when we eventually build a UI for OpenWiki, it’ll likely help with exposing content to humans (better sectioning, discovery, preview details, etc) Anyone else started using it yet? Super curious to hear more

RT Alex Olsen We need to talk more about model <> harness <> task fit This 🧵 is your excuse to learn why this is so important

> For me, agentic performance on spreadsheet tasks is the whole of Shortcut existence. So I'm going to beat competitors by obsessing over the harness. custom harness is the only way you will beat the labs agentic experience heres how to build one: https://www.langchain.com/blog/how-to-build-a-custom-agent-harness

RT Brace OpenWiki 0.1.2 is now out, with a ton of changes! Notable updates: - 🧑‍💻 QOL: `openwiki --init` and `--update` run Code Brain by default (most popular use case!) - 🔙 Specific prompting to document backlog tasks within wiki files - 🥳 8 new contributors, with this release including 11 total unique contributors!! Try it out today, and put up a PR if you have a new feature, fix or improvement! https://github.com/langchain-ai/openwiki

if you want to learn more about why coding harnesses are not going to be the optimal harness for science: https://www.youtube.com/watch?v=RjpTrffSMjE

RT Nicholas Larus-Stone Re @LatchBio has been on a tear recently. It should be obvious to everyone why coding harnesses are not going to be the optimal harness for science, but it was nice to see someone put numbers to it

RT Himanshu | AI Engineer Started contributing to LangChain's OpenWiki. Today I got a comment from @hwchase17 "Thanks for contributing!" Small moment, big motivation. Time to keep shipping and contributing more to open source. 🚀 #opensource #LangChain #AI #buildinpublic

RT Josh Rosen Don’t let them tell you it’s all about the models. It’s just as much about the infra.

LangSmith does this for you - in the cloud: LangSmith sandboxes & LangSmith deployments - any model: LangChain integrates with 100s - harness: deep agents - tracing: langsmith observability - recursive improvement: langsmith engine

📕LLM Wiki Webinar with @BraceSproul @devstein64 @jeffreyhuber is now on YouTube! Two of my favorite insights from this webinar: "Wiki as a cache" - @devstein64: basically, the purpose of a wiki is to keep things that are commonly looked up or accessed more top of mind/readily available. What goes in the wiki should be updated and organized with that in mind "Wiki is a set of hyperlinked pages" - @jeffreyhuber. The evolution of memory (in my view) has gone from: 1. single string 2. file 3. set of files in a directory (wiki) as you think about scaling up - as files grow, the file structure matters less, and you need links between pages. The internet is a set of hyperlinked pages - so that definitely scales! Watch it here: https://www.youtube.com/watch?v=Lsut4TCfygw

RT Brace OpenWiki now supports ChatGPT login!! Use your subscription instead of API credits when running OpenWiki Thank you @Topzsixx for the contribution to add support! Try it out today 👇

RT Ankush Gola I use OpenSWE multiple times a day directly from slack. Makes it super easy to go from conversation -> code

RT Thomas Reaves 1/ I watched LangChain’s webinar on LLM wikis and agent memory and wrote up a working paper from my own small agent-fleet experiments: Soft-Cache: A Human-Supervisable Coherence Protocol for Persistent Agent Fleets https://github.com/treaves-GSD/soft-cache-agent-indexing

great overview of general purpose brain!

RT Brace Another big langchain week

🚀langchain launches this week: all about open source models and memory! First: open source models. We partnered with @NVIDIAAI to launch a NemoClaw DeepAgents blueprint. This pairs Deep Agents (our open source, model agnostic harness) with Nemotron 3 ultra (powerful OSS model) and OpenShell (enterprise ready run time). Blog: https://www.langchain.com/blog/langchain-and-nvidia-launch-the-nemoclaw-deep-agents-blueprint Second: memory. LLM wikis continue to intrigue us. @BraceSproul @colifran_ released a new version of OpenWiki focused on "personal brains" - creating wikis from gmail, internet, etc: https://github.com/langchain-ai/openwiki open models and memory pair well together as well. They both contribute towards companies and enterprises owning their whole stack - from the model level all the way up to the context level

RT Brace OpenSWE is one of our most widely used agents throughout the company. Since July 1st it's been tagged over 700 times in Slack! This doesn't even count the reviewer agent, tagging it in GitHub, or tagging it from Linear tickets It's 100% open source too

RT Colin Francis great viz by @BraceSproul to put this in perspective. the really important part of this is that OpenWiki is additive!

Sierra isn't the first to build this - Ramp, Stripe, CoinBase also have If you want an open source version - check out OpenSWE: https://github.com/langchain-ai/open-swe We use it internally (mostly for coding). Model agnostic, fully OSS but integrates seamlessly with LangSmith for o11y

RT Brace OpenWiki general purpose memory is meant to be complementary to codex/claude code memory: it's proactive & ambient, meaning it'll automatically go out into your world (via connections like gmail, x, notion, etc), discover what you're working on or interested in, and remember it for future reference this paired with reactive memory (codex/claude memory) leads to an incredibly powerful agent that knows everything about you to have truly comprehensive memory, you can't stick with one or the other since both types are useful and important

love this framing of memory as "proactive"

RT Colin Francis agent memory has always been reactive. OpenWiki makes it proactive. connect to sources, tell it what you care about, and your agent hits the ground running . what we're building is really exciting! try it out and get involved, it's open source! 👇 https://github.com/langchain-ai/openwiki

RT Evangelos Kostopoulos Thank you @hwchase17 @BraceSproul @devstein64 @jeffreyhuber for the LLM Wikis session today, I absolutely loved it! Really grateful you keep these sessions open and honest about what actually works in the industry. The idea I'm taking home is eventual correctness, where the wiki self-corrects over time until it converges on the truth. That's exactly what I'm chasing with my agents too. Thanks @LangChain for hosting, can't wait for the next one!

RT LangChain "Whoever Jerry is, he was excellent." That's a customer talking about an agent. @PodiumHQ's Walker Ward sat down with our COO @j_schottenstein to share how LangGraph + LangSmith helped his team take their agents from prototype to production.

RT Colin Francis OpenWiki Brains v0.1.0 is out 🎉 personally really excited about personal brains! these maintain context about you and what you're working on, interested in, etc. webinar for all things wiki happening now: https://events.langchain.com/webinar/llm-wiki/

big update to openwiki - it now supports two modes: - code brain: create a wiki for a code base - personal brain: create a wiki for more general purpose tasks (email, web, etc) webinar at 11am PST (5 minutes!) where we will be talking about wikis! https://events.langchain.com/webinar/llm-wiki/

RT Brace http://x.com/i/article/2075248391349768192

this is going to be a fun one! clay is one of the most ai-forward thinking companies i know of lots to learn here if you're in NYC

langsmith for coding agents

Setup evals with… terraform??

RT Adam Łucek You've heard of Infrastructure as Code- but agent evals can now ride your existing Terraform setup! I've been using the new LangSmith Terraform provider to auto-provision online evals + monitoring alerts for my agents, and think you should too. Here's how 🧵

RT Nick Hollon spent a ton of time working on this!! so happy to see the results come out. very impressive model from @NVIDIAAI that we fit our harness to so we could get better performance!

RT @cvmilo00: tbf this is one of the best repos on github, specially if you are REALLY into AI

Bigger update coming tmrw, but a bunch of small fixes today!

i want to build a more opinionated/detailed coding experience (for prompt optimization, eval set creation, some more things) what is the best way to package it all up?

Guess what we added someone else to this webinar: @jeffreyhuber, ceo and cofounder of chroma What does a vector database think about the concept of “wikis”? Come find out! It’s turning into a party

Congrats to prime intellect! Love partnering with them on LangChain labs work

Great opportunity!

RT Niko Come help us scale @harvey’s model training team. If you’re interested in bringing frontier agent research into the Harvey product and working with: - @baseten to scale up RL to 80M+ token virtual datarooms - @PrimeIntellect to create structured agent training environments from unstructured legal data - @FireworksAI_HQ to navigate the quality <> cost Pareto frontier with inference-time routing and advisor models - @LangChain & LangChain Labs to build efficient verifiers and close the observability <> training feedback loop - @appliedcompute to post-train open weight models and high-volume agents for end-to-end legal tasks - @EngramLab to create an entire synthetic law firm and firm knowledge memory systems for better / more efficient open-world search - @trajectorylabs & @NVIDIAAI to shape the frontier of continual learning and sovereign AI for high-stakes domains - @mercor_ai & @SnorkelAI to build out Legal Agent Bench and other benchmarks across legal and other verticals and other projects like this, then this is the role for you. Apply here: https://www.harvey.ai/company/careers/d78083d9-a203-4ae4-b4b9-454d65df3702

RT Gabe Pereyra We are hiring for @Harvey’s model training team. This team will help Harvey expand from the application layer into the model layer and from legal into high end knowledge work more broadly. We are hiring AI researchers of all seniority, particularly those with experience post-training frontier or open source models. Our program is centered around large-scale model training, synthetic data generation, long horizon reinforcement learning, and rigorous evaluation in real world deployments. We are scaling-pilled and believe that nothing beats the combination of larger models and better training data. We’ve been able to generate incredibly realistic legal environments and validated that this allows us to post-train open source models to achieve frontier performance with agents. We plan to scale up these data generation and training efforts significantly across legal to start, and eventually other verticals. As a researcher, you will have access to thousands of GPUs and unique training data from our product and customer relationships. Your research will inform Harvey’s product strategy and power AI used for some of the most economically and societally impactful work in the world.

Love partnering with baseten to make sure everyone can use open weight models in deep agents

Really excited to partner with @nvidia on the NemoClaw Deep Agents Blueprint Deep Agents is a fully open source agent harness that we are tuning to make perform incredibly well with open models

RT Alex Olsen I cannot put into words how stoked I am on this It's not an understatement to say that this blueprint could represent the future of enterprise inference

gpt-5.6 sol isnt the only thinking launching thursday we're also releasing a big update to OpenWiki (auto create wikis of code bases... and more?) we're also going live with a webinar to talk all things wiki: https://events.langchain.com/webinar/llm-wiki/ (openwiki: https://github.com/langchain-ai/openwiki)

agents for ad spend

RT Anuj Patel The most important part of an AI agent isn't the LLM. It's what surrounds it. 👇 The first image made it click for me. While exploring LangChain Deep Agents, I ran the same prompt. Same model. Almost the same answer The harness decides how the agent thinks and acts That's why two agents using the same LLM can behave very differently My takeaway: As models become similar, better harnesses will matter more

RT Mykhailo Chalyi Re @hwchase17 ATIF proposes to have steps as a field in JSON document. This means that more or less longer sessions especially with some binnary content would result in unprocessable huge JSON file. Sounds like no starter. My best guess community should adopt PI session format, or codex.

is anyone standardizing on ATIF as a format for agent traces? or a different format? or is it still wild wild west, roll your own

deepagents is our newest open source project - an open source, model agnostic agent harness this is maybe the most important academy course we've launched

llm wikis are a glimpse of the future of what agent memory looks like this blog i wrote resonated with a lot of folks will be discussing this (as well as some updates ive made to my beliefs since this!) this thursday with @BraceSproul @devstein64 https://events.langchain.com/webinar/llm-wiki/

How a local-first personal AI agent uses orchestration, memory, tools, LangGraph, background workflows, and child agents

RT 𓁟 SYD 🛸 I wrote an article on the core Agent architecture for Row-Bot. How it uses a @LangChain LangGraph agent at its core for the main agent, background workflows and child agents. Each is a full LangGraph ReAct agent Here is the full article @hwchase17 : https://x.com/SydSachar/status/2074499930236821550

Good blog from viv on how improving agents (via rl, harness Eng, anything) boils down to a data mining problem over traces

RT Viv http://x.com/i/article/2073138207986966528

RT Coframe We drove 410% conversion lift on @Replit's enterprise funnel, 2x-ing the number of demo requests. All in a matter of months. "Coframe makes you feel like you have a team of 100 people. There's nothing on the market like it." ⧉ x ⠕

RT Taylor Dolezal Come join us on Thursday! It'll be a great session you'll want to add to your MEMORY.md 😄

devin has a hot take that wikis are NOT the right abstraction should be a fun webinar!!

RT Devin Stein does your agent need a wiki? should humans and agent use the same wiki? should you have one wiki or many wikis? if you're wiki-curious, join @BraceSproul @hwchase17 and I on Thursday! you don't want to miss it @dosu_ai 🤝 @LangChain

RT Brace We're going to be discussing LLM memory & wiki's this week with @hwchase17 and @devstein64! Don't miss it

RT Brace big things happened over the weekend!

RT LangChain Last minute webinar this Thursday Will cover all things related to "LLM Wikis"

🚨Emergency webinar: LLM Wikis and how to give your agent memory LLM Wikis are so hot right now - OpenWiki by @BraceSproul up to nearly 7k GitHub stars in less than a week I'll be chatting with Brace and @devstein64 about Wikis this Thursday: https://events.langchain.com/webinar/llm-wiki/

RT Caspar Broekhuizen Soon most agents for work will run in the cloud and communication with them will happen almost exclusively in your existing work channels (Slack, Teams, ...) This hasn't become the norm yet because local-first platforms are much easier to engineer. Running on the user's machine avoids the harder product surface: hosted VMs, identity, sharing, permissions, long-running state But these mechanics are necessary for agents to become shared team software so will become table stakes

RT LangChain Product Manager @BenTannyhill on why LangSmith Engine routes trace investigation through screener + verifier sub agents instead of letting the main agent read everything.

RT Julia Schottenstein new favorite episode of Max Agency with @bentannyhill and @hwchase17 ! They sit down to talk about how we built Engine, our agent for agent engineering, at @LangChain. So much to love about this episode - spans everything from the who builds agents, how we did it, and tools we used to get it done. if you can't tell, I've never been more excited about a product at LangChain! listen here 👇 https://www.youtube.com/watch?si=wp2AKnhbVnFSCzb8&v=YqjR4vQwbTc&feature=youtu.be https://podcasts.apple.com/nz/podcast/the-best-ai-agents-are-secretly-teams-ben-tannyhill/id1891551672?i=1000775182821