Agent harness

An agent harness is everything in an agent that isn't the model: the system prompt and instruction files, the tools and their descriptions, how context is assembled and trimmed, hooks and middleware, any sandbox or subagents, and the loop that calls the model, runs what it asks for, and decides when to stop. Claude Code and Codex CLI are harnesses, and so is the code around the model in a team's own support agent; the same model can sit inside any of them.

It exists because a model on its own only predicts the next message. It can't run a command, keep state between sessions, or notice it's repeating the same failed edit. The harness turns those predictions into actions, with memory and limits around them.

On every turn the harness decides what the model sees, which tools it can reach, and what happens to its output, so it shapes behavior alongside the weights. It's also the part of an agent a team keeps editing: a prompt line, a new tool, a changed default.

A harness change can make an agent worse with the model untouched. In EvoHarnessBench, a 2026 benchmark that grows an agent's harness stage by stage, adding tools, skills, or specialist agents, with nothing else changed, could lower results on tasks the agent had already solved. Harness engineering is a young discipline, so writers still draw a harness's boundary differently: some include observability and evaluation, others only the runtime around the loop.