Unconstrained agents captivated Twitter. Custom ones are what ship.

Pat Grady and Sonya Huang of Sequoia Capital’s Training Data introduce Harrison Chase as the product visionary who first connected language models with tools and actions. LangChain, the company he founded, is the most popular agent-building framework in the space. The hosts want a definition that can survive the hype, because everyone is talking about agents and almost no two people mean the same thing.

Chase treats the word as early and contested, then offers a clean cut. In a traditional RAG chain the steps are known ahead of time: generate a search query, retrieve documents, write an answer, return it. An agentic system puts the model at the center and lets it choose. Sometimes it searches. Sometimes it answers immediately. Sometimes it searches three times and then replies. The model is deciding the control flow of the application.

Tool use and memory get talked about as if they *were* the definition. Chase sees them as companions. If the model is choosing what to do, tools are how it acts, and memory is how it knows what it already did. Grady asks whether agents are really about decision-making or about taking action. Chase says the two are bound together: the hard part of acting is deciding which action is right, and a system around the model then executes that choice and feeds the result back.

The distinction from a chain is a spectrum, not a cliff. At the simple end, a router classifies which path to take. At the other end sit fully autonomous agents. In between is a lot of gray. Chase generally agrees with Sequoia’s bet that the industry is moving from copilots to agents, because a copilot still needs a human in the loop and that caps how much work an external system can do. The paradox is obvious: the more you let the agent run, the more leverage you get, and the more it can go off the rails. Striking that balance is the design problem ahead.

Asked where LangChain sits in that ecosystem, Chase points at the middle of the spectrum. Fully autonomous prototypes are simple to stand up and they go off the rails a lot. Pure chains are too rigid for the jobs people actually want done. The company is trying to be the orchestration layer that lets builders create agents that are more constrained than AutoGPT and more flexible than a hardcoded DAG.

That focus is a product of the company’s own history. Early LangChain mixed chains with a single AgentExecutor class and then kept adding knobs to that class. Users wanted far more control than one class could give them. The response was LangGraph, an extension aimed at customizable agents that live between chains and autonomy — loops, persistence, and human-in-the-loop included.

Grady takes Chase back to March 2023, when AutoGPT and BabyAGI captured Twitter. AutoGPT became one of the most popular GitHub projects ever, which Chase treats as a peak of the hype cycle. He felt a lull from late summer 2023 into the start of 2024. Then realistic systems started to come online: Elastic’s assistant and agent in production, Klarna’s customer-support bot, Cognition’s Devin, and Sierra.

The AutoGPT-style architecture failed for a blunt reason. It was very general and very unconstrained, which is why it captured imaginations. In practice, the work people want to automate is specific. They want rules followed and a particular way of doing things. The agents Chase sees in the wild are custom cognitive architectures: directed enough that you would not just write ordinary code, flexible enough that the model still chooses inside a lane. That takes more engineering, more trial and error, and more time, which is why it did not exist a year earlier.

A cognitive architecture, in his usage, is just the system architecture of an LLM application: from user input to user output, what data and model calls flow, whether there are branches or cycles, which steps are algorithmic and which are left to the model. As teams try to get agents into production, those graphs get more bespoke. There are checks up front, a handful of allowed next steps, substeps, and loops back. He calls it *cognitive* because the power of the models is reasoning, and the graph is a human mental model for a task encoded into software.

Grady hears two things at once: the graphs are bespoke, and they are fairly hardcoded. Is that the destination, or a stopgap until elegant default architectures show up? Chase takes the question seriously. At one extreme, if models become reliable enough at planning, the best design might be a for-loop: call the model, decide, act, repeat, with every constraint stuffed into the prompt. He expects planning and reasoning to improve. He does not think that loop will be the best way to build for a long time.

The objections are efficiency and reliability. If you already know that A must follow B, put them in order; do not spend a model call re-deriving the sequence. And LLMs are non-deterministic. In an enterprise setting you want comfort that the step that is always supposed to come first actually always comes first. Building these systems should get easier, and the graphs may get less ornate. The shape of production, though, stays custom.

Then he offers what he flags as a hot take. Plot cognitive architectures on two axes: simple versus complicated, generic versus custom. A for-loop is simple and generic. Production systems are custom and complicated. The remaining quadrant — complicated but generic, the elaborate planning step, the reflection loop, tree of thoughts — he expects to shrink, because generic planning and reflection will be trained into the models. What will not be trained in is domain-specific reasoning: the control loops that belong to a particular business. He is bullish on both ends of that remaining spectrum.

Huang cites Princeton’s SWE-agent paper: coding agents now solve 12.5% of GitHub issues, up from 3.8% with RAG, which is still not intern-level. Chase splits the question. SWE-agent is a relatively general agent, asked to work across many repositories. Something like Vercel’s v0 is far more reliable than 12.5%, and Elastic has shipped multiple agents into production, including a new one around RSA. General agents are still hard. Longer context, better planning, and better reasoning will help them. The ones that already work are custom.

Chase had shared a Jeff Bezos line with the hosts: early-twentieth-century breweries generated their own electricity, and the lesson for founders was to focus on what makes the beer taste better. Huang asks the current version of the question. Does owning your cognitive architecture actually improve the product, or should you cede that layer to the model lab and spend your time on UI? Chase splits it the same way he split the quadrants.

A generic cognitive architecture does not make the beer better. Model providers will work on general planning, and off-the-shelf loops will keep improving. A custom one does: encoding how your support team thinks, how an internal process actually runs, or the best way you know to write a particular kind of code. If applications are going to *do the work*, that bespoke logic is a large part of what you are selling. UX, UI, and distribution still matter. The line to draw is general versus custom.

Grady then asks Sequoia founder Don Valentine’s famous question: so what? If autonomous agents work flawlessly, how is life different? Chase’s picture is that a lot of repetitive industrial work gets automated, and people spend their time deciding what the agents should do and building on the outputs. You could bootstrap a company as a CEO with a marketing agent and a sales agent, outsourcing functions you would otherwise hire for, and keep the strategic and product thinking. The point is to free people for what they want to do and are good at.

Two categories already have traction. Customer support is having an impact today; LangChain itself needs support, and hiring agents for it would be powerful. Coding is the second, less mature, and for Chase more personal: his mother has an idea for a website and cannot code it. An agent that closed that gap would let her stay on the idea. Grady hears Dylan Field’s theme from Figma — closing the gap from dream to reality. Chase adds that being a builder in the age of generative AI no longer means you have to be an engineer or hire one. Intelligence is getting cheap enough that a much larger set of people can make things.

Chase walks the architecture timeline as a sequence of failures. Early models could not even reason well about the first step, so chain-of-thought prompting gave them room to think; that habit then got trained in and became default. Shunyu Yao’s ReAct paper, which Chase calls the first cognitive architecture for agents, added an explicit reasoning step before each action and ran it in a loop. As models internalized that reasoning, the written “thought” became less necessary. ReAct-style agents today often just use function calling. What remains from the paper is the loop.

Once the loop runs, two problems show up: planning, and noticing you are done. Models struggle to form a long-term plan, track progress inside the loop, and avoid declaring victory after step one of five. The architectural fix is to make the plan explicit, then walk it step by step. Reflection is the cousin: after a search or an answer, ask whether the job was actually done well instead of returning garbage. Planning and reflection are the popular *generic* architectures. Chase expects them to be trained into models. The custom graphs tied to business logic will not be.

At Sequoia’s AI Ascent, Chase had argued that UX can influence the effectiveness of the architecture — usually treated as opposite ends of the stack. Models still mess up, which is why chat won: you see the stream, you correct it, you ask a follow-up. Chat also keeps the human in the loop, which is copilot-shaped and caps leverage. The interesting patterns try to keep a human available without making them drive. Devin’s transparent list of steps lets you rewind to step three, edit the decision, and continue. An inbox pattern lets ten agents run in the background and ping you only when they are stuck.

Review is the other family. Research agents such as GPT Researcher can write a first draft; you leave comments in a batch, or you collaborate in a Google Docs style where the agent fixes one comment while you write the next. Underneath all of it is learning: if a person has to give the same correction a hundred times, the UX has failed, and the system architecture has to remember. Chase says every one of these patterns is still early, and that this is a large share of what his team thinks about.

Grady notes that Chase is unusually present in the developer community, which means he sees problems LangChain will not productize. Chase lists the obvious ones: the model layer, vector databases, and fine-tuning infrastructure, where companies like Fireworks already play. LangChain will help with data curation. It will not train foundation models or run the fine-tuning cluster. Further out, if an economy of agents actually arrives, the missing pieces look like identity, permissioning, and payments — including the inversion where an agent pays a human to do something. He thinks that market is still too immature to pick winners.

Fine-tuning and prompting are not substitutes. A more custom cognitive architecture narrows what each node is asked to do, and narrow tasks are where fine-tuning shines. The open-source LangChain project standardized interfaces across models, vector stores, and tools, plus higher-level starters for RAG and SQL Q&A. LangGraph is for the cyclic case: persistence so you can resume, run in the background, and deploy long-running human-in-the-loop apps. LangSmith, in the works since the company started, is observability and testing for systems whose center is a non-deterministic model, plus a prompt hub and a human annotation queue.

On observability, Chase thinks something new is required. Datadog is strong at monitoring individual traces; multi-step LLM apps need cross-trace insight, and developers stare at single traces precisely because non-determinism makes each failure a puzzle. Testing is even more different. Software tests pass or fail; the eval people actually trust is LMSYS-style pairwise comparison in an arena. Pass rates may never hit 100%, so the job is to track the trend and catch regressions. And humans have to look, because software can check that 2 equals 2 as well as you can, while the quality of a model’s answer often cannot.

In the lightning round he admires OpenAI’s execution over the prior year and a half, including Sam Altman and Logan, and he singles out researcher David Dohan, whose early model-cascades papers shaped how Chase thinks. He also admires Zuckerberg and Meta for the Llama open-source bet. As a product-centric CEO he looks to Brian Chesky, whom he heard at Sequoia Base Camp. His advice to founders is one line: just build. GPT-5 will make some of what you did irrelevant, and you will learn too much along the way to wait. Grady’s coda: at Sequoia’s first AI Ascent in early 2023, Chase sat in the audience pushing code all day — the day OpenAI shipped plugins.