deep read: @hwchase17
2026-02-20
founder of langchain and langsmith. the most influential and most criticized figure in the llm tooling space.
4 articles, jan–feb 2026. all focused on the same thesis: observability is the foundational layer for agent systems. written during the langsmith agent builder launch. these are the most honest and most self-serving articles simultaneously — the diagnoses are genuinely correct, and the destination is always langsmith.
context
harrison chase built langchain in late 2022 — the right framework at the right moment. 90k+ github stars. then the backlash came. over-abstracted, poorly documented, constantly breaking APIs. "remove langchain" threads became a genre. his response: langchain the framework took a backseat. langsmith — the observability platform — became the real product. he's no longer selling the framework. he's selling the insight layer on top of whatever framework you use.
the good
the observability thesis is correct. you can't predict agent behavior from reading the code. agents make decisions at runtime based on context, instructions, and model behavior that changes across versions. the only way to understand what an agent does is to watch it work. this is genuinely true and genuinely important.
the filesystem-shaped memory is a practical design. instead of flat key-value memory, langsmith's agent builder uses a file tree: user_info.md, project_context.md, conversation_history/. the agent decides what to remember and where to file it. this is more natural than most memory architectures and scales better than a single growing context window.
"why you can't predict agent behavior" is his strongest piece. the core argument — code tells you what an agent can do, not what it will do — is the kind of fundamental insight that most agent builders learn the hard way. the gap between capability and behavior is where most production failures live.
the bad
4 articles is a thin corpus. every article ends at langsmith. the sandbox pattern piece is a product tutorial wearing an essay's clothes. the memory system design is interesting but underdeveloped — no forgetting mechanism, no consolidation, no conflict resolution. the writing is careful to the point of being cautious. you never catch chase saying something he might regret, which also means you never catch him saying something surprising.
verdict
read "why you can't predict agent behavior." it's the essay that most clearly articulates the observability gap in agent systems. the rest is product marketing with good production values.