Context Engineering Our Way to Long-Horizon Agents: LangChain’s Harrison Chase
LangChain’s Harrison Chase argues context engineering — not just better models — is the core unlock behind long-horizon agents finally working.
- Claude Code’s rise is attributed largely to harness quality, not just model quality — Terminal Bench 2 shows Claude Code is not top of leaderboard.
- Harrison noticed the long-horizon agent inflection point in June–July 2025 when Claude Code, Deep Research, and Manus all converged on the same core loop architecture.
- Every serious long-horizon agent harness today uses file system access; Chase calls himself ‘file system pilled’ and says it is non-negotiable for context management.
- Agent memory creates compounding moat: Chase’s own email agent became worse when moved to a new system that lost accumulated memories, blocking his migration.
- Traces, not code, are the new source of truth for debugging agents — teams share LangSmith trace links instead of GitHub links when something breaks.
- Junior developers and younger founders outperform senior engineers at agent building because they lack preconceived notions about deterministic software.
- ‘Sleep-time compute’ — agents reviewing their own daily traces overnight and updating their own instruction files — is the next memory frontier Chase is building toward.
- Browser use is not ready; code execution in sandboxes is the reliable general-purpose tool, with coding agents potentially becoming the default general-purpose agent architecture.
2026-01-21 · Watch on YouTube
