The Anatomy of an Agent Harness

A nice description of the anatomy of a harness with the memorable formula Agent = Model + Harness

Agent = Model + Harness; the harness is all code, config, and logic outside the model that turns raw model intelligence into a functional agent.

Harnesses provide durable state via filesystems, git for versioning, bash/code execution for general-purpose autonomy, and sandboxes for safe, scalable tool use.

Memory uses filesystem files like AGENTS.md for continual learning; web search and tools overcome knowledge cutoffs.

Context rot is fought with compaction, tool output offloading, and skills for progressive disclosure to preserve reasoning quality.

Long-horizon work relies on filesystem state, Ralph Loops for continuation, planning, and self-verification loops to enable autonomous complex tasks.