Agent hooks

Instructions in a prompt are requests; the model usually follows them but can forget or misjudge. Hooks are code, so they run every time. You register a command against an event, and the harness executes it when that event fires.

Claude Code, for example, lets you attach hooks to events such as before a tool runs, after a tool runs, when you submit a prompt, and when the agent stops. A hook that runs before a tool call can inspect the requested command and block it, while one that runs after a file edit can format the code or run a linter.

Hooks are a good place for rules that must never be skipped: blocking edits to protected files, refusing dangerous shell commands, logging every action, or running tests before the agent declares a task done. Keep them fast and predictable, because they run on every matching event.