LangChain

I joined Michael Kennedy on Talk Python To Me with Reuven Lerner, Calvin Hendryx-Parker, and Patrick Decker to talk about "The State of Python in 2026." So much of it came back to development with agents! Harnesses, observability, and why ~90% of my coding happens in Slack now. https://lnkd.in/gdPNF_J7

Harrison Chase joined 20VC with Harry Stebbings for a wide-ranging conversation on where agents are headed. Harrison talks through why neutral agent harnesses still matter, why he’s bullish on vertical agents, what most people get wrong about open models, and so much more. https://lnkd.in/giwsiDwK

"It's underrated how hard it is to build agents!" Great conversation between Harrison Chase and Harry Stebbings on 20VC. An agent = model + context + harness, and you can improve an agent by tweaking any of the three. A few takeaways: 1. The highest ROI thing you can do is start by optimizing context (then the harness, then the weights). This is where domain knowledge lives, and that's incredibly valuable. 2. The story of continual learning is: run an agent, collect traces, aggregate learnings, fix something, measure the fix, loop! The fix can land in any of the three layers. Those fixes all look different, but the cycle is the same. 3. Memory is still massively unsolved, but also massively important. The hardest part is distilling learnings from traces, which is incredibly domain specific. 4. When you build agents, you're probably worried about a few things: accuracy, cost, latency, and security. We've seen these themes confirmed year after year in our dev surveys. 5. When we think about reducing cost, the pipeline looks like adding visibility into spend, putting controls in place, and only then pushing on agent optimizations. This is how we reduced our coding agent spend at @LangChain. 6. Model routing, which is great for cost controls, is most effective in the harness (where the context needed for a routing decision lives). 7. Open models are used for more narrow tasks right now (small parts of the agent loop, or very specific domains with post-training), but as they improve they'll power more core agent loops. 8. We're looking at a future with both horizontal agents and vertical agents. Horizontal agents will use the vertical agents, and infra needs to change to support this load (agents can do background work at a rate that humans cannot). https://lnkd.in/gqpP7vkS

Numan built an eval-first AI Health Coach, Nu, using LangGraph and LangSmith. Nu doesn’t diagnose or prescribe. It coaches, and it must not cross the medical device line at any point. Their solution is an eval-first approach: every eval runs through the same graph patients talk to, scored per category, with simulated patients played by LLM personas. Asif Laldin, Marcus Wadge-Dale, Shubhangi Hora, and Konstantinos Georgakilas from the Numan team put together a blog post that dives into how they simulate and score these conversations, how they hold the regulatory line, how they can ship sub-agents in weeks and migrate frontier models in just a day, and what this architecture looks like. Give it a read below ⤵️ https://lnkd.in/ekAdn2XF

Interested in a career at LangChain? RSVP for our SDR recruiting social on October 27th. 🤝 Meet LangChain's sales team and connect with fellow sales professionals 👥 Enjoy food, drinks, and networking 🏢 Hear firsthand about the SDR role, team culture, and career growth 💬 Participate in an optional speed interview for the chance to fast-track your application Interested in attending? RSVP while spaces last. https://luma.com/ubhnxwnu

Scientific research agents can do more than retrieve information. They can synthesize studies, work across research sources, execute analysis, and help teams explore complex biomedical questions. The challenge is making those workflows reliable enough to trust. Join our EMEA Deep Life Sci Workshop: Agent Engineering for Life Sciences for a hands-on look at how to build, trace, evaluate, and improve research agents for complex life sciences workflows. 🌍 Register for the EMEA session: https://lnkd.in/gdSeyp25

✅ Everything we shipped at Interrupt NYC and London ✅ Open source updates ✅ Upcoming events & livestreams Catch up on everything LangChain in our latest newsletter. https://lnkd.in/ekpFCFBC

Harness² from Google Research offers techniques for recursive self-improvement (RSI) on the agent harness, with a fixed model. They call it "harnessing," but you might already know this process as "hill climbing" or "continual learning". It works at two levels: 1. Task level: analyze a trajectory, propose a change, validate the change, and commit it if validated. 2. Harness level: learn which changes are effective, so task-level optimizations get better over time. This pattern is very familiar to us at LangChain! It comes down to two fundamental units: 1. The trace: the trajectory of each run. Harness² requires each edit to be justified by evidence in a trace. 2. The eval: whether a change actually made the run better. https://lnkd.in/g-HgpM_B

Your agent traces can become training data Agent observability systems like LangSmith ingest massive volumes of agent traces. Compared to traditional traces, they are a) larger and much more information-dense b) serve a different purpose: continually improving the agent, not just helping you find where your system is slow or broken Two ways to continually improve your agents with traces: harness engineering (Engine) and post-training. For post-training, traces become fuel for: - SFT examples: successful prod trajectories can teach a base model - Tasks: real prod agent inputs become eval / RL tasks - Verifiers: trace feedback from users + other sources can help define success criteria for agent outcomes - Tool snapshotting: traces plus source code can reconstruct tools and state to build agent environments

I have always considered Miles Grimshaw to be one of the smartest investors in venture. I remember Sept 2023, Miles came to our recording studio in London having just joined the Benchmark partnership. At the time, he had done one deal; a small company called LangChain.  Miles raved about the founder Harrison Chase. His excitement for the company was palpable.  Today the company is worth $1.2BN and serves 35% of the Fortune 500.  The other day I sat down with Harrison and he allowed me to add my notes from the chat below.  1. What Percent of Developer Salary Do You Spend on Coding Tools and How Has It Changed? LangChain spends 30% of its engineering salary on coding tools, a share that has fallen over the last two months even as token usage has gone up. The team tracks coding agent usage in LangSmith, sets budgets the VP of engineering can raise, and controls the harness. Its in-house cloud coding agent uses prompt caching and routing to cut costs. 2. What Will Happen to the Inference Market? Harrison says he underestimated the demand for inference and for coding agents, both far bigger markets than he expected three or four years ago. Echoing Eric Vishria's line that "Everything just works," he believes any inference provider will work, and he does not expect demand to drop anytime soon. 3. Why Harnesses Need to Be Neutral Companies want to be able to switch models, but the harness can be tightly tied to a model, and context today is intricately linked to the harness. The harness connects the model to data, APIs and memory. If memory is built in a way that is linked to the harness, and the harness is linked to the model, the company ends up stuck. 4. Why the Memory Layer Will Become More and More Important and How Instinct Crushed It Memory is rare in agents today but will be incredibly common in the next few years. Instinct and Muse have nailed personalization and learning, so the product feels like it is for the individual user, through its tone or proactive follow-ups. Harrison still sees memory as super early. 5. Will We Have One Horizontal Agent or Tons of Vertical Agents in the Consumer Realm? Consumers will have both, with horizontal agents interacting with vertical ones. Most apps should have a vertical agent because they can learn more about how a user interacts with them than a general agent can. Wrapping an app's many APIs in a single natural language agent also gives horizontal agents a better experience than raw APIs. 6. What Do Large Enterprises Care Most About Today? Cost has risen as a concern for large enterprises, though it still ranks below accuracy and above latency. Security is a bigger concern for large enterprises but still sits below accuracy and cost. 7. Why Is Benchmark the Best Fund? Benchmark invested in LangChain when it was just Harrison and Ankush, and the founders treated the firm as a third co-founder. As the company grew from two people to around 400, Benchmark has been its most consistent partner.