Type: url
Original: https://blog.langchain.com/the-anatomy-of-an-agent-harness/
Status: generated
Summary
A March 2026 LangChain blog post that defines an agent harness as every piece of code, configuration and execution logic that is not the model itself — “Agent = Model + Harness” — and derives a harness’s core components by working backwards from what a model cannot do out of the box. It covers filesystems and git for durable state, bash and code execution as a general-purpose tool, sandboxes with default tooling for safe execution and self-verification, memory files and web search for knowledge beyond the weights, compaction, tool-call offloading and Skills against context rot, and Ralph Loops, planning and self-verification for long-horizon work. It closes on the co-evolution of model post-training and harness design, arguing that the best harness for a task is not necessarily the one a model was trained with.