Harness
Everything around the model that makes it an agent: the loop, tools, permissions, memory, context management.
The same model behaves very differently in different harnesses. Claude Code, Codex and Cursor are harnesses: they decide which tools exist, what needs approval, how files are read, when context is compacted. When an agent fails, the harness is often the fix, not the model.
Related terms
- Agentic loop: The cycle of model call, tool call and result that an agent repeats until it stops.
- Hooks: Scripts the harness runs automatically at fixed points, such as before a tool call or after an edit.
- Context engineering: Deciding what goes into the model’s context, in what order and when, across a whole task.
