The agent's hooks

Hooks are commands the coding agent runs before and after each tool call, and when it is about to stop. A hook sees each call as it happens, so it can record the call or refuse it.

Finding it

They are set in the coding agent's settings: in Claude Code, the hooks section of .claude/settings.json in the project, or of the user's own settings. A person approves them, since a hook runs with that person's rights; /hooks in Claude Code lists the active ones.

Using it

• Record: a PostToolUse hook appends each call and its result to a log, for patterns read later.
• Refuse: a PreToolUse hook that exits with code 2 blocks the call, and what it writes to standard error goes back to the agent as the reason.
• Before the agent stops: a Stop hook can run the checks and, when they fail, exit with code 2 to keep the agent working, with the failure as its reason.
• Quick and narrow: a hook runs on every matching call, so it is fast and matches only the tools it is about.

Examples: Claude Code hooks (PreToolUse, PostToolUse, Stop), and the equivalents in other coding agents.