Harrison Chase

Self healing deployments is the future

RT Gocha Berulava this is why we're building the loop the other way around too deploy → monitor → pipe logs back to the agent via mcp → agent fixes the code → redeploy observability is step one. self-healing deployments is step two https://x.com/hwchase17/status/2027094058330656814 Original tweet: https://x.com/gochaberulava/status/2027982438438174923

Reliability for agents is not just driven by engineers PMs and smes are very involved

RT Nikhil Ramesh Awesome content from start to finish - the last point about reliability now being a cross functional problem involving Engg, PM and SMEs is something thats often overlooked. Original tweet: https://x.com/nikhil2197/status/2027946221403574554

RT Kevin Simback 🍷 Re @TyRobben @hwchase17 I disagree, what I took away from this article is that there are companies like @LangChain that are rapidly building the tools needed to deploy an agentic workforce at scale Original tweet: https://x.com/KSimback/status/2027851841833238721

RT Muktesh Couldn't agree more. Original tweet: https://x.com/mukteshkrmishra/status/2027831920604942766

RT Ty Robben Read this article and you’ll l realize anyone saying ai is going to eliminate a huge swath of white collar jobs this year is way over their skis It’s the old overestimating what happens in 5 years under estimating what happens in 10 just condensed to a 1yr/5yr (maybe 3yr) time frame instead Original tweet: https://x.com/TyRobben/status/2027770366006071478

RT Ankur Kumar As usual, a must read for Agent Observability in Prod - particularly liked these points 👇 1️⃣ Human evals play the role 2️⃣ LLM based continuous evals based on domain context 3️⃣ Annotation based evals 4️⃣ Insights Agent to find what you are looking for in prod Original tweet: https://x.com/ankurkumarz/status/2027554278827823267

RT Thom Agent monitor for production Original tweet: https://x.com/ThomGuillemard/status/2027530945763414067

RT Tadeo Donegana Braunschweig Built a store management agent with @LangChain deepagents Shoutout to @hwchase17 @Vtrivedy10 @sydneyrunkle @masondrxy for the posts that got me started Original tweet: https://x.com/tadeodonegana/status/2027479756157960618

RT Tadeo Donegana Braunschweig http://x.com/i/article/2027107697410121728 Original tweet: https://x.com/tadeodonegana/status/2027478696286658778

RT Darshj.AI Re @hwchase17's "context isolation through sub-agent delegation" is underrated. In production we spawn ephemeral agents with scoped context windows — they burn tokens in isolation, compress results, and die. Orchestrator never bloats. The real unlock: treating agent lifetime as a context management primitive. Original tweet: https://x.com/thedarshanjoshi/status/2027475984362934613

RT kevin Re @hwchase17 Non determinism + high permutation of user input leads to entirely different approach to qa, tracing in real time in prod is best! Original tweet: https://x.com/kleffew94/status/2027445220053533016

RT arvindkampli Re @hwchase17 or- you can’t improve if you don’t measure Original tweet: https://x.com/ArvindKampli/status/2027434444869800151

RT Steve Jarrett Brilliant blog post by @hwchase17 about the differences with debugging agents versus traditional software. Great to see the @LangChain team continuing to work in such an open way on these state-of-the-art challenges. Original tweet: https://x.com/stevejarrett/status/2027413898396098797

RT Contextually | Cue Re @hwchase17 This. And you can't personalize what you don't remember. The agents that win long-term are the ones that carry context forward - not just per-session, but across every interaction. Contextually | Cue coming soon get it first: https://contextually.me/ Original tweet: https://x.com/ContextuallyAI/status/2027406577095901457

RT Abhishek I love that @LangChain @hwchase17 never shy away from sharing some of the best practical Agent Engineering wisdom. Apart from @AnthropicAI, I haven't seen many companies doing this a regular cadence. Original tweet: https://x.com/abhi__katiyar/status/2027386278225027214

RT Joshua Ebner AI agents and AI generated software are going to create a new explosion of data. Get ready. Original tweet: https://x.com/JoshuaEbner/status/2027380068092113016

You can’t evaluate what you don’t store

RT Codefly Great article on something I think a lot about. The big challenge here is a massive shift in data volume and architecture. We’re moving from tracing millisecond latencies to storing and indexing entire multi-turn conversations. We’re seeing a 100x increase in the data we need to collect just to understand why a decision was made. You can’t evaluate what you don't store. Original tweet: https://x.com/code_fly/status/2027360474086482301

RT Jianchu Xu (JC) Xu Very inspiring! Original tweet: https://x.com/jc_jianchu_xu/status/2027323406614708471

RT Sam Crowder When I used to work on databases, I knew what my app would do once I had written unit/integration tests. with agents, I don't know what my app will do until it's deployed and doing things! Original tweet: https://x.com/samecrowder/status/2027186521888137356

RT LangChain JS `@langchain/[email protected]` just landed! 🚀 Three big additions: tool streaming, tool progress UI, and state overwrite support. Here's the breakdown ↓ Original tweet: https://x.com/LangChain_JS/status/2027106070896992756

http://x.com/i/article/2027092716484726784

RT Ash Lewis Next week I'm speaking alongside Sid Bidasaria (co-creator of Claude Code) and @hwchase17 from @LangChain at Coding Agents by @mlopscommunity My talk: why agents need small language models. 📅 March 3rd 📍 Computer History Museum, Mountain View 🔗 https://luma.com/codingagents Original tweet: https://x.com/ash_csx/status/2027081301786865690

RT LangChain ⏳ Last chance to submit a talk for Interrupt (May 13-14, SF) Call for papers closes Friday 2/27. Share what you've learned building and scaling agents in production. Submit your talk 👉 https://interrupt.langchain.com Original tweet: https://x.com/LangChain/status/2027077451113771455

RT ByteDance Open Source DeerFlow 2.0 looks forward to your feedback at GitHub http://github.com/bytedance/deer-flow @hwchase17 @LangChain Original tweet: https://x.com/ByteDanceOSS/status/2026859269594009674

RT Henry Li 20k+ stars on Github. Now we're going further. DeerFlow 2.0 — rebuilt from scratch on LangGraph 1.0. From Deep Research to Long-horizon super agent harness, with planning, long-term memory, file system, and SKILLS! https://github.com/bytedance/deer-flow @ByteDanceOSS @hwchase17 @LangChain Original tweet: https://x.com/henry19840301/status/2026847782456406331

RT Sydney Runkle Come learn about the bleeding edge of agent development! I'll be in SF all next week - shoot me a DM if you're building w/ deepagents and want to chat! Original tweet: https://x.com/sydneyrunkle/status/2026739703995679164

RT LangChain JS We just shipped deepagents-acp — an open-source TypeScript library that turns any @LangChain agent into an ACP server. One package. Any model. Every IDE. @zeddotdev, @teamcity, Neovim, Emacs — your agent works everywhere, no vendor lock-in. https://www.npmjs.com/package/deepagents-acp Original tweet: https://x.com/LangChain_JS/status/2026684368538960285

langsmith can trace claude code! so when you think claude code is nerfed... you can set up some observability to back that up

RT Jai Bhagat Re @bcherny @trq212 OK, so this is INSANE. 1189 calls to Claude. 100% nerfed down to Sonnet 4.5 in the last 30 days despite Claude Max. I'm so happy I have LangSmith for observability. There could be a bug on how this is reported. But right now, this is really bad... cc: @hwchase17 @Vtrivedy10 Original tweet: https://x.com/ChaiWithJai/status/2026446654753190324

RT Sam Crowder Half of the Fortune 10 use LangSmith for observing and evaluating their agents! And 35% of the F500 use LangChain products. Original tweet: https://x.com/samecrowder/status/2026381556307374450

RT Brace 🤖 Agent Builder now supports queuing messages ⏰ Started a task in Agent Builder but forgot to include extra instructions? Now, you can submit as many followups as you'd like, and they'll be automatically added to a queue. Try it out for free today: https://www.langchain.com/langsmith/agent-builder Original tweet: https://x.com/BraceSproul/status/2026333549348237782

RT Christian Bromann 🪄 In the next version of http://zeitzeuge.dev (v0.10.0): FCP rendering diagnostics. It now records your page load via @ChromeDevTools screencast, analyzes it frame-by-frame, and pinpoint exactly what's delaying 1st paint, with a visual filmstrip embedded right in your report! Original tweet: https://x.com/bromann/status/2026316184849711592

RT LangChain JS In the next version of deepagents (v1.8.1) 🪄 👉 custom namespace isolation for StoreBackend (user/org-scoped storage) 👉 critical OOM fix in conversation history offloading 👉 improved subagent ToolMessage handling 👉 examples for handling streaming scenarios Original tweet: https://x.com/LangChain_JS/status/2026311155199750293

Many teams treat evals as a last-mile check. http://monday.com Service made them a Day 0 requirement for their AI service agents. Using LangSmith, the monday service team has been able to: 🔷Achieve 8.7x faster evaluation feedback loops (from 162 seconds to 18 seconds). 🔷Get comprehensive testing across hundreds of examples in minutes instead of hours 🔷Gain agent observability with real-time, end-to-end quality monitoring on production traces Read more on their eval-driven development here: https://blog.langchain.com/customers-monday/

RT Christian Bromann 🔥 Hot take: API reference docs shouldn't be built for humans anymore. So we built http://reference.langchain.com for agents first — MCP server, llms.txt, machine-readable content negotiation on every page. The pretty UI is just a courtesy for the remaining humans 🤖 Original tweet: https://x.com/bromann/status/2026026309173031361

RT LangChain JS Introducing http://reference.langchain.com — a unified API reference for every @LangChain package across Python, JavaScript, Go & Java. 100+ packages. Updated daily. Version history for every symbol. Built with AI-first mindset: 👉 A built-in MCP server so your coding agent can look up any LangChain symbol 👉 llms.txt for instant LLM context 👉 Content negotiation — every page serves markdown or HTML depending on who's asking 👉 http://chat.langchain.com built in — ask "how do I use ChatOpenAI with streaming?" and get an answer right there Original tweet: https://x.com/LangChain_JS/status/2026025198903034104

RT Christian Bromann New in http://zeitzeuge.dev v0.9.0: native support for @bunjavascript and @nodejs test runners 💫🐢 ZeitZeuge now profiles your test suites beyond Vitest — just add a preload script (Bun) or reporter (Node.js) and get AI-powered analysis of your slowest tests. No config files, no wrappers. Original tweet: https://x.com/bromann/status/2026021831652880569

RT Sam Crowder Google ADK tracing for LangSmith! Folks wonder all the time if LangSmith only works for our frameworks like LangChain and LangGraph. Given our naming conventions, I can't really blame them. but it couldn't be further from the truth! Original tweet: https://x.com/samecrowder/status/2026014126397596083

RT LangChain 🔎 We shipped native tracing for Google ADK! See how easy it is to get started observing your ADK agents in LangSmith with just a few clicks. LangSmith works natively with over 25 frameworks and providers, and not to mention OpenTelemetry! 🔥 Docs 👉 https://docs.langchain.com/langsmith/trace-with-google-adk Original tweet: https://x.com/LangChain/status/2026013755193327687

RT LangChain 🚀 New updates to the LangSmith filtering experience when viewing Traces 🚀 It’s now easier to: • Apply filters and edit filters • See active filters at a glance Happy tracing! Sign up for LangSmith ➡️ https://smith.langchain.com/?utm_medium=social&utm_source=twitter&utm_campaign=q1-2026_langsmith-fh_aw Original tweet: https://x.com/LangChain/status/2025994394143219803

RT LangChain How Exa built a production-ready deep research agent with LangSmith and LangGraph 👀 Exa, known for their fast, high-quality search API, has a deep research agent that delivers structured answers on the web -- no matter how complex the query. Powered by LangGraph, they've built a multi-agent system. For Exa, one of the most critical LangSmith features was observability, especially around token usage. "The observability – understanding the token usage – that LangSmith provided was really important. It was also super easy to set up." – Mark Pekala, Software Engineer at Exa. This visibility into token consumption, caching rates, and reasoning token usage proved essential for informing Exa's production pricing models and ensuring cost-effective performance at scale. Original tweet: https://x.com/LangChain/status/2025744946494345570

RT Git Maxd Went looking for what "only changed the harness" actually means in code One PR caught my eye 👀 Inline prompt strings -> a 239-line system_prompt.md 4-step loop: Understand -> Build -> Test -> Verify The harness engineering thesis is in a single PR Original tweet: https://x.com/GitMaxd/status/2025725851695042595

RT LangChain 🌟 LangSmith Insights Agent 🌟 Use LangSmith Insights to group traces and find emergent usage patterns of your agents 🔎 Now with the ability to set a schedule and run recurring jobs! Docs: https://docs.langchain.com/langsmith/insights Original tweet: https://x.com/LangChain/status/2025612841819025834

RT LangChain OSS LangChain Community Spotlight: agent-debugger 🐛🔍 A terminal debugger for LangGraph and LangChain agents featuring semantic breakpoints—pause execution based on agent behavior patterns, not just code lines. Unified visibility into agent decision-making and Python execution. 🔗 https://github.com/dkondo/agent-tackle-box Original tweet: https://x.com/LangChain_OSS/status/2025601480443789728

RT LangChain http://x.com/i/article/2025365243061567489 Original tweet: https://x.com/LangChain/status/2025366346973007956

RT LangChain OSS LangChain Community Spotlight: langchain-agent-skills 🚀 Production-ready LangGraph patterns with AI assistants. Lubu Labs built a research agent in 6 days vs 2-3 weeks—orchestrator-worker architectures, state reducers, selective retries, and LangSmith debugging. Full case study: https://www.lubulabs.com/ai-blog/langchain-agent-skills-real-world-example Original tweet: https://x.com/LangChain_OSS/status/2025269293177614779

RT Git Maxd People asked why Deep Agents was faster in that side-by-side video I made for @Vtrivedy10’s epic article So I traced both with LangSmith to find out Turns out Claude Code searches for every file before reading it. Even when the filename is explicitly prompted🔍 Traces below👇 Original tweet: https://x.com/GitMaxd/status/2025006165265187233

RT Jason Yuan 🕊️❤️🦜 Original tweet: https://x.com/jasonyuan/status/2024914168714068390

RT Jason Yuan very excited to be cohosting a machine learning <> dinner dinner next week with @hwchase17 and the folks at @LangChain we’re trying to bring together the most creative first principles thinkers so if that’s you pls sign up! Original tweet: https://x.com/jasonyuan/status/2024913901931106812

I'm hosting a small dinner with @jasonyuan next week focused on the intersection of design and ML In SF, very limited spots. If you are a designer interested in AI, or an ML engineer who cares deeply about design - we'd love to host you! sign up: https://luma.com/d5dw9vzf

RT Abejith Agent Observability by @hwchase17 at @LangChain "You don’t know what your agent will do until your users use it" "Production is where you discover what to test for offline. The traces automatically become your eval dataset" Original tweet: https://x.com/Abejith/status/2024623464691511656

RT Brace Want to learn how agent builder manages memories? We’ve written an article detailing exactly how: Original tweet: https://x.com/BraceSproul/status/2024618080350150816

RT LangChain OSS RLM in Deep Agents! Original tweet: https://x.com/LangChain_OSS/status/2024590527211901397

RT LangChain JS LangChain now has a first-party OpenRouter integration → available in both Python and Typescript. Access 300+ models from OpenAI, Anthropic, Google, and more through a single interface. Tool calling, structured output, and streaming all work out of the box. One API key, works with every model, and without juggling SDKs Get started in one line: > uv add langchain-openrouter > pnpm install @langchain/openrouter Original tweet: https://x.com/LangChain_JS/status/2024582319613603868

many people are saying

RT Hunter Lovell you wouldn’t believe how many times we hear this exact same story Literally everything we ship (open source or otherwise) is a manifestation of the opinions and lessons learned that come from working with a countless number of people who run into these same problems. a lot of people naturally choose the path of writing their own tool calling loops since its easy to understand on the tin, but the value sell of an agent framework these days isn't just that: it's a way to not re-learn all the same lessons we've learned when you need to put agents in front of real people Original tweet: https://x.com/huntlovell/status/2024573277348516013

RT vinny I LOVE @LangChain DEEP AGENTS @hwchase17 please keep it up Original tweet: https://x.com/vinicius2prg/status/2024557386002821616

RT LangChain LangSmith Agent Builder uses memory to improve with feedback. Three practical ways to get the most out of memory: → Tell your agent to remember what works → Use skills to give it specialized context when needed → Edit its instructions directly when that's faster Full walkthrough: https://blog.langchain.com/how-to-use-memory-in-agent-builder/?utm_medium=social&utm_source=twitter&utm_campaign=q1-2026_ab-philosophy_aw Original tweet: https://x.com/LangChain/status/2024556612455977005

RT LangChain 🚀 Introducing: LangSmith for Startups We are launching LangSmith for Startups to give early-stage teams the tooling and community they need to iterate faster and win bigger. Observe, evaluate, and deploy your agents -- now with free credits to help you get started. 💸 Our Scale program offers: ✅$10,000 in LangSmith credits ✅Exclusive programming: Technical sessions directly with the LangChain team ✅Community: High-level networking with peer founders ✅Visibility: Product feature & recruiting spotlights on our social platforms We’re proud to partner with the world’s leading investors to bring this to life: @sequoia , @ycombinator , @AmplifyPartners, @southpkcommons, @MayfieldFund, @NEA, @BessemerVP, @firstround, @BasisSet, @Lux_Capital, @kleinerperkins, @AforeVC, @Accel, @khoslaventures, and @AudaciousHQ. How to join: If you're a startup backed by one of our partner VCs and Series A or earlier stage, you qualify! Join today 👉 https://www.langchain.com/startups Original tweet: https://x.com/LangChain/status/2024545770100211931

RT LangChain 🚀 New updates to the LangSmith filtering experience when viewing Traces! It’s now easier to: • Apply filters and edit filters • See active filters at a glance Happy tracing! Original tweet: https://x.com/LangChain/status/2024540855256961325

RT Christian Bromann Re @jarredsumner 👋 I build a tool called ZeitZeuge that captures v8 profiles as part of your unit tests and analysis it for you using Claude: https://zeitzeuge.dev/ (see announcement post: https://x.com/LangChain_JS/status/2024515544788140134) .. wonder your thoughts 🤔 Original tweet: https://x.com/bromann/status/2024521213897486560

RT LangChain JS What if your test suite could tell you exactly what's slow 🤔 and write the fix? 🤯 ZeitZeuge is a Deep Agents app that captures V8 CPU profiles from your @vitest_dev tests, hands them to 4 specialized subagents, and outputs code-level perf fixes. We ran it on LangGraph. It found 13 bugs code review missed 🙌 Read how evals took it from 60% detection to 100%, and how LangSmith traces debugged the agent itself 👇 Original tweet: https://x.com/LangChain_JS/status/2024515961274106009

RT LangChain JS http://x.com/i/article/2024510466278772738 Original tweet: https://x.com/LangChain_JS/status/2024515544788140134

RT Mason Daugherty deepagents-cli supports sandboxes out of the box! choose between @daytonaio @modal @RunloopDev or roll your own with our sandbox protocol our CLI is: - open source - (actually) model agnostic - 5th on terminal bench 2! docs here: https://docs.langchain.com/oss/python/deepagents/cli/overview#use-remote-sandboxes Original tweet: https://x.com/masondrxy/status/2024348794872008816

When building datasets to evaluate agents, do you use synthetic data generation?

RT Tristán Sepúlveda 🤖 Building a rental pre-screening AI on WhatsApp | LangGraph, GPT-4o & Fraud Detection 🛡️ | Python Full-stack learner Original tweet: https://x.com/TristanSBuilds/status/2024249529902440511

RT Mason Daugherty deepagents-cli==0.0.23 is out! 📸 - drag & drop image attachments for chat input - expanded local context details + bash-based impl for sandbox support - skill deletion command (thanks @breath57!) - startup time reduced in half - bugfixes + UX enhancements all around https://github.com/langchain-ai/deepagents/releases/tag/deepagents-cli%3D%3D0.0.23 Original tweet: https://x.com/masondrxy/status/2024233787924865409

RT Sam Crowder we just shipped baseline experiments in LangSmith. an easy way to compare new model changes, prompt tweaks, or parameter tuning against an existing source of truth. Original tweet: https://x.com/samecrowder/status/2024209495476621535

RT LangChain We just shipped Baseline Experiments 🚀 You can now pin any experiment as your baseline in LangSmith. This allows you track performance deltas, anchor your results, and quickly identify improvements or regressions in an experiment list. Docs: https://docs.langchain.com/langsmith/analyze-an-experiment#set-a-baseline-in-the-experiments-view Original tweet: https://x.com/LangChain/status/2024208662936650152

RT Viv this is a really cool role! the team is really tuned into the frontier of agent research & evals, lots to learn and figure out together + we get to openly share with an awesome community of builders :) related note, Harrison is VERY involved in making great content (go read his blogs) and shipping with the team + the folks are just really nice great to place to grow and try things, DM him! Original tweet: https://x.com/Vtrivedy10/status/2024194539594940433

the agent space is still SO EARLY we spend a lot of time teaching people the art of possible and best practices around agents, evals, observability we're hiring someone to join our dev rel team to help with this effort apply here (or dm me!): https://jobs.ashbyhq.com/langchain/0b5c9efa-aec8-451d-a24a-a54865c46924

RT LangChain 🚀 We just shipped a major update to LangSmith Agent Builder: • New agent chat:&nbsp;One always-available agent with access to all your workspace tools • Chat → Agent:&nbsp;Turn any conversation into a specialized agent with one click • File uploads:&nbsp;Attach files directly to Agent Builder • Tool registry:&nbsp;Add, authenticate, and manage your tools in one place Try it now: https://smith.langchain.com/agents?skipOnboarding=true/?utm_medium=social&utm_source=twitter&utm_campaign=q1-2026_agent-builder-chat-launch_aw Learn more: https://blog.langchain.com/new-in-agent-builder-all-new-agent-chat-file-uploads-tool-registry/?utm_medium=social&utm_source=twitter&utm_campaign=q1-2026_agent-builder-chat-launch_aw Original tweet: https://x.com/LangChain/status/2024180357457989887

RT Git Maxd Side by side example Same model (claude-opus-4-6). Same task. Two different agent harnesses @LangChain Deep Agents CLI: 9s Claude Code: 16s The harness IS the performance. 1.7× difference, zero model changes Original tweet: https://x.com/GitMaxd/status/2024137171217871106

RT Raphael Fraysse This was one of the key learnings of my past week in SF! Great article Original tweet: https://x.com/la1nra/status/2024056279619260604

RT Sonya Huang 🐥 hell yea @zeddotdev @LangChain Original tweet: https://x.com/sonyatweetybird/status/2023981740784775452

RT Viv Astasia summed it up perfectly :) 🙏🏽 It’s valuable looking into the harness as a vector for agent improvement + unlocking tasks that a different/worse harness simply wouldn’t allow (ex: long horizon agentic coding w/o iterative self verification) if anyone’s interested, we’re also actively thinking about simply picking the “right model” per sub-task in a multi-model harness, ideally driven autonomously rather than by us ex: Codex for planning, Gemini for multimodal understanding, all working together would love to hear any thoughts on this and if there’s use cases we can help you unlock! Original tweet: https://x.com/Vtrivedy10/status/2023943900642046293

RT Astasia Myers Great post! 🚨 Don’t change the model. Change the harness. The unlock? • Forced self-verification • Trace-driven debugging • Context injection • Smarter reasoning budgets Harness engineering > model switching Original tweet: https://x.com/AstasiaMyers/status/2023939654005964814

RT Viv Building a System for Agent Self Verification: Some folks asked about the ONE thing for good harness building from the blog there’s no one thing, but I suggest starting with: “how to build a self-verification loop into the harness via tests” this is the best bang for buck for long horizon autonomous coding agents. The agent needs mechanisms to: - verify correctness of code sub-pieces with existing + generated tests - integrate via non-negotiable connector interfaces to other pieces of code (integration contracts). These contracts help guide correctness of new code - plan for recovery and replanning when the initial build fails - use hooks to reorganize context after each build phase in the file system. Use codemaps that point to files, don’t dump all context into one file, it’ll slopify over time without external context management there’s more but thinking through what your agent needs to stay on track over many turns is a good team exercise for more content: @dexhorthy and the humanlayer team do a great job talking about verification, testing, and backpressure @ryancarson’s recent Code Factory blog is also a great working example of this vision Original tweet: https://x.com/Vtrivedy10/status/2023918215936553355

RT Atai Barkai Excited to be participating in @LangChain's Interrupt 2026! If you're building with agents, this is the place to be at. Come talk with me and the @CopilotKit team about building fullstack agentic applications with generative UI! Signups are open now 👇 https://x.com/LangChain/status/2022005146201600469?s=20 Original tweet: https://x.com/ataiiam/status/2023886462375326096

RT Riya this is so good man Original tweet: https://x.com/riyajaiinn/status/2023872409091403810

RT letsbuildmore Re @hwchase17 Sdk is great. I am using deep agents in a product. I have both Claude agents and deep agents running with a toggle switch. Found deep agents to be super fast in reply compared to Claude agents (could be due to it having to proxy over Litellm for Gemini model) Original tweet: https://x.com/letsbuildmore/status/2023854461987471587

Very cool community project! 🪲agent-debugger agent-debugger is a terminal debugger for LangGraph/LangChain agents. It combines agent-level visibility (state, messages, tool calls, store snapshots, and semantic breakpoints) with Python-level debugging (line breakpoints, stepping, stack, and locals) in one Textual UI. Repo: https://github.com/dkondo/agent-tackle-box?tab=readme-ov-file

RT Viv I’m a fan of this “Systems Engineering” approach to Agent Development Every new piece we add/remove is an exercise in integrating systems because we can’t add things to an agent in isolation. The pieces affect each other Evals are a bridge to measure both specific capabilities like (ex: testing skill use) or overall perf in a domain (“did the latest harness changes improve the agent on coding and/or accidentally reduce some other capability”) Evals are hard to come up with as well, it’s a classic blank slate problem We use Traces to guide our Eval generation process. Some questions we think about: - What behavior from the traces do we want to make sure we keep doing or stop doing in the future? - How do we break down this trace into small smoke tests for actions Eval generation is a pretty creative exercise. Everyone does evals differently but broadly mapping small specific unit/smoke tests and some tasks in a domain (coding) is a decent start Original tweet: https://x.com/Vtrivedy10/status/2023836267729797323

if youve got an agent running in production, it can actually be quite hard to understand how users are actually using your agent but this info is super valuable! only by knowing how people are using your agent can you actually improve the experience Insights solves that

RT Mason Daugherty Deep Agents via ACP! Original tweet: https://x.com/masondrxy/status/2023816052753469448

Great deep dive into harness engineering

ACP is my dark-horse contender for the next protocol to explode You can now easily spin up an ACP server for any deep agent!

RT Jacob Lee Lately, I've gotten super excited about @zeddotdev's open-source Agent Client Protocol. It's a fantastic way to pass a coding agent the exact context you're looking at in an IDE. That's why as a pat leave side project I built my own ACP client to replace my Claude Code usage! It's a few hundred lines of Python hooked up to a @LangChain Deep Agent with filesystem and shell access. It works great out of the box, but the real power comes from customizing the internal agent to use whatever prompts, models, tools, skills etc. you want. Plus, you get full observability via LangSmith! I've been using it the past few weeks to great effect. See the links in the replies to try it yourself ⤵️ Original tweet: https://x.com/Hacubu/status/2023804529314243012

RT Christian Bromann Deep agent architectures spin up multiple subagents in parallel. But until now, streaming treated everything as a flat message list. 👉 No visibility. No progress. No trust. 👈 We just shipped first-class subagent streaming in @LangChain_JS 🤯 Here’s how to build real-time multi-agent UIs 👇 https://www.youtube.com/watch?v=hrEWqm8JA-w Original tweet: https://x.com/bromann/status/2023799237805568385

RT Itamar Friedman Introducing Qodo v2.1 🎉 Starting today, Qodo will automatically learn, surface, track and enforce your custom Rules. This is a pivotal moment in code quality. Some aspects of code quality are generic, we’ll agree upon that. But some are tech stack-specific, dev org-specific, repo-specific, and even subjective and opinionated, if you like. For these reasons, forward-thinking teams are manually collecting Rules and Skills. It is a tedious task, but a critical one if you care about the quality of your code and software development. To have a complete set of Rules and Skills that fit your standards (or even exceed them!), you need to spend time; a lot of time. We've seen complete, meaningful sets ranging from dozens to hundreds in a single repo, and even into the high thousands across a complete codebase. Moreover, these Rules are dynamic. They change. They change since libraries change, performance requirements change, policies change, design patterns change, etc. So these Rules must be learned and they MUST adapt. We also deserve to know if the Rules and Skills we set are actually meaningful and useful? Are they actually considered by AI tools? Do code review tools like Qodo find issues accordingly? How many? Do developers follow these Rules? At Qodo we care a lot about enabling professional dev teams that want to harness AI to code fast, AND still care about their code quality, because they understand what waits for them along the line later if they don't care. With Qodo v2.1, installed with a few clicks, you can get all of the above. Qodo automatically learns your code base, your PR history and related metadata, then offers just-in-time Rules suggestions that you can accept/decline, edit, and start tracking and enforcing. Qodo will even read your http://Claude.md, http://Agents.md, http://Rules.md, etc.. to extract Rules from there. When Qodo finds a rule violation, it will include a link to the relevant rule so you can see it...

RT Mason Daugherty Long-running agents face the same problem: conversations exceed models' context limits, so older tokens must be dropped. Many tools handle this well for users; for instance, tree-based navigation that lets you browse and restore any point in conversation history via a --resume. But what about the agent itself? When the agent needs a detail that was lost in summarization, most systems require user intervention. The user must navigate the history, find the relevant context, and re-inject it. We took a different approach with Deep Agents: 1. Before summarizing, we offload history to `/conversation_history/{thread_id}.md` 2. Include the file path in the summary message 3. The agent can read this file with standard tools whenever it needs more detail The agent knows where to look and has the tools to read. It's a small architectural choice with big implications for agent autonomy. Original tweet: https://x.com/masondrxy/status/2023607596578398459

RT LangChain 👌 Tracing in LangSmith is as easy as copy/paste 📊 Get started in seconds with Claude Agent SDK, OpenAI, LangChain, Vercel AI SDK, and 20+ other frameworks. Pick your stack, copy the code, start debugging. Docs: https://docs.langchain.com/langsmith/integrations Sign up for LangSmith: https://smith.langchain.com/?utm_medium=social&utm_source=twitter&utm_campaign=q1-2026_langsmith-fh_aw Original tweet: https://x.com/LangChain/status/2023532973086159283

RT Tom Loverro Had a great chat with @hwchase17 at @LangChain's first all-company offsite. We got into something I think every scaling founder needs to hear. An astute LangChain employee asked: “Isn't shipping quickly at odds with craftsmanship and product quality?” My counterintuitive answer: No—speed to customer hands improves quality. The sooner users touch your product, the sooner you know if you're building the right thing and the sooner you find bugs. Quality comes from contact with reality, not longer internal bake times. Playing it safe can actually kill product quality at growing companies. Teams keep iterating on fixes and small features internally, but nothing ships for months and the really big, new roadmap item gets bogged down in red tape and death-by-committee. We saw this firsthand at @coinbase . After the 2017 crypto boom, product velocity stalled in 2018 as headcount grew quickly and the company focused on infrastructure (eg, keeping the site up and ready for scale) and navigating regulatory ambiguity (eg, avoiding all risk). It took @brian_armstrong and @balajis cutting through the red tape to get back to shipping. Whenever they were told something would take 6 months or “is not possible” they kept asking “Why?” until they found the root cause. Turned out customers cared more about having new tokens to trade than the 6th nine of uptime. For Series B founders, the #1 thing to protect as you scale is product velocity on legit new product (not just iterating existing features). Growth brings bureaucracy. Your job is to cut through it. Original tweet: https://x.com/tomloverro/status/2023428407552557566

🦜Ciana Parrot Self-hosted AI assistant with multi-channel support, scheduled tasks, and extensible skills Kind of like OpenClaw but on top of deepagents! Fun project from someone in the community: https://github.com/emanueleielo/ciana-parrot

RT LangChain How Exa built a production-ready deep research agent with LangSmith and LangGraph 👀 Exa, known for their fast, high-quality search API, has a deep research agent that delivers structured answers on the web -- no matter how complex the query. Powered by LangGraph, they've built a multi-agent system. For Exa, one of the most critical LangSmith features was observability, especially around token usage. "The observability – understanding the token usage – that LangSmith provided was really important. It was also super easy to set up." – Mark Pekala, Software Engineer at Exa. This visibility into token consumption, caching rates, and reasoning token usage proved essential for informing Exa's production pricing models and ensuring cost-effective performance at scale. Read about how they built their agent here: https://blog.langchain.com/exa/?utm_medium=social&utm_source=twitter&utm_campaign=q1-2026_langsmith-fh_aw Original tweet: https://x.com/LangChain/status/2022732667305730397

RT Aaron Levie File systems are an agent’s natural work environment. The ability to process and create unstructured data allow agents to bring automation to most areas of knowledge work. Now you can easily integrate Box as a cloud filesystem into deepagents from Langchain. Stay tuned for more. Original tweet: https://x.com/levie/status/2022375298097111160

RT LangChain JS [email protected] + [email protected] — here's what shipped this week: - Summarization middleware rework (no more full state rewrites) - Hardened skills validation + annotations - `wrapModelCall` middleware can now return `Command` - More accurate token counting with tools Here's the full breakdown ↓ Original tweet: https://x.com/LangChain_JS/status/2022048105626317167