Mental Model

A hook is an event contract between the agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose runtime and deterministic code. The contract has five parts:

1. Event — the lifecycle moment that triggers evaluation.
2. Payload — the action, context, or result supplied to the handler.
3. Decision — observe, modify, allow, deny, or request another model pass.
4. Failure semantics — what happens on timeout, crash, malformed output, or missing configuration.
5. Scope — the local, project, managed, IDE, or cloud surface where the hook exists.

Products can expose similarly named events with different control power. A pre-tool event that may deny an action is materially different from a notification that runs before the same action.

Product Semantics

Claude Code documents several handler types and a broad lifecycle. Codex exposes local hooks but describes them as guardrails rather than complete enforcement. Copilot exposes hooks across named agent surfaces, while documented timeout behavior can fail open.

The right comparison is therefore not “does it have hooks?” It is “can this exact event, on this exact surface, produce this exact decision under failure?”

Trust Boundary

Hooks execute close to the agent loop and may share its machine, credentials, or configuration authority. They are excellent for rapid feedback and telemetry. They are insufficient as the only release gate when the same actor can edit, disable, bypass, or time out the handler.