Agent Event Protocol (AEP) is an open draft standard for normalizing AI coding agent activity and hook control responses.
For the current draft version, see VERSION. Breaking changes are expected before 1.0.
AEP defines:
- a canonical event envelope for prompts, model output, tool/action activity, sessions, notifications, and orchestration;
- a canonical hook response shape for control decisions such as allow, deny, ask, defer, block, continue, stop, and abstain;
- a shared action lifecycle that keeps immutable events separate from control responses;
- a declarative mapping format for existing agent-native hook payloads.
The goal is interoperability: producers can emit one shape, and consumers can build policy, approval, observability, and audit systems without per-agent event branches.
This is an early public draft. Use GitHub to participate:
- Open issues for unclear wording, missing fields, lifecycle gaps, or provider-specific cases that should generalize.
- Open discussions for design questions or broader feedback.
- Open pull requests for typo fixes, examples, mapping drafts, or proposed spec edits after discussion.
Repository: https://github.com/agentapprove/agent-event-protocol
We especially want feedback from agent runtime providers and teams building hooks, approval systems, policy engines, observability tools, and agent dashboards. Provider representatives can open an issue or discussion in the repo, or contact [email protected] if a direct conversation would be easier.
This repository contains formal AEP draft content: the spec, integration guidance, mapping drafts, examples, license, changelog, and contribution notes.
It does not include product-specific adapters, unpublished hook code, runtime code, service configuration, or deployment details.