Context Engineering Our Way to Long-Horizon Agents: LangChain’s Harrison Chase

LangChain’s Harrison Chase argues context engineering — not better models — is the core unlock for long-horizon agents, and traces are the new source of truth replacing code.

• Claude Code’s popularity is primarily a harness achievement, not a model achievement — and it’s not even at the top of Terminal Bench 2.
• Anthropic trains on explicit file-editing tools; OpenAI trains heavily on Bash — harnesses must be tuned per model family, not just per model.
• Context compaction is still fully manual in most harnesses; Anthropic has an experimental mode letting the model decide when to compact, but adoption is low.
• Agent logic lives partly in the model, not the code — you can’t read the codebase to understand behavior, you must read the traces (LangSmith is built around this shift).
• Memory is a real moat: Chase’s personal email agent became noticeably worse when migrated to Agent Builder because it lost accumulated memory, and he still hasn’t fully switched.
• “Sleep-time compute” (term from Letta): agents reviewing their own daily traces overnight and rewriting their own instructions is the near-term self-improvement pattern Chase is most bullish on — over RL.
• Junior/younger engineers are consistently outperforming senior engineers at agent development because they lack preconceived notions about deterministic software.
• Traversal (Sequoia-backed) and Klara (K) cited as production examples of long-horizon AI SRE and customer support agents respectively.
• File system access is non-negotiable for long-horizon agents; browser use is not ready; code execution is ~90% necessary; virtual file systems (Postgres-backed) work for context management but can’t run code.
• The inflection point was June–July 2025 when Claude Code, Deep Research, and Manus all converged on the same architecture: LLM in a loop with context engineering — not new algorithms.

Guests: Harrison Chase (LangChain cofounder), hosted by Sonya Huang and Pat Grady (Sequoia) · 2026-01-21 · Watch on YouTube