20VC with Harry Stebbings · Oct 17, 2026 · 1h 3m · 52m saved
With Harrison Chase, Co-Founder and CEO of LangChain
LangChain spends roughly 30% of what it pays its engineers on AI coding tools, Harrison Chase, co-founder and CEO of LangChain, told Harry Stebbings on 20VC.
Much of the debate about AI agents is about which model lab comes out ahead. Harrison's answer is that the winners are the people who can tell whether an agent did its job well, and those are the experts in each field.
"If value is derived through verification, I think the people who win are the domain experts... I'm massively bullish on vertical agents for different domains for this reason." — Harrison Chase
Harrison started LangChain in 2022 with his co-founder Ankush, when the company was two people. Its framework for building agents spread fast enough that more than a hundred competing frameworks followed it, and Benchmark and Sequoia backed it early.
The full interview is covered here so you can skip it. 1 hour 3 minutes of audio, 11 minutes of reading.
Here are the 11 arguments that matter.
Key Takeaways
Domain experts who can write evals will capture the value in AI agents, which is why Harrison is bullish on vertical agents
As models improve, LangChain has to erase big parts of its own product
Agent memory is unsolved, and LangChain killed its own commercial memory product for being too general
Model routing is hard to build and not a commodity, because a new frontier model can wipe out its advantage
LangChain spends roughly 30% of its engineering salary bill on coding tools
One misconfigured plugin cost about $40K in a week
Enterprise fear of Chinese open-weight models is "largely unfounded" when they are self-hosted or run through a provider
Founders should take a valuation discount to work with Benchmark or Sequoia
1. Agents Misbehave by Design
The core difficulty with agents, in Harrison's account, is that models are non-deterministic black boxes. The range of possible inputs is effectively infinite, and a small change in an input can produce a very different output.
Harrison Chase says agents chain so many steps that small errors compound
"The hard part about building agents is they don't behave how you want them to behave." — Harrison Chase
An agent is a sequence of model calls, each one interacting with outside tools. Every link in that chain is another place for an unexpected input to send the result somewhere new.
"Agents are not just one model call. There's a... series of model calls and they're interacting with other tools." — Harrison Chase
