An open specification for the lifecycle hooks every AI agent runtime should expose so that operators, auditors, and regulators can record what an agent did, when, why, and with what reasoning.
The specification is Apache-2.0. Stewardship sits with Agentic Thinking Limited (UK) under the commitments set out in GOVERNANCE.md. Adopters benefit from a perpetual no-relicence pledge and a working-group governance model.
The first reference implementation is HookBus. Other compliant implementations are explicitly invited and will be tested against the same conformance suite.
EU AI Act Article 12 (record-keeping obligations) becomes enforceable on 2 August 2026 for high-risk AI systems falling under Annex III categories. Every enterprise deploying agentic AI must demonstrate auditable, transcript-grade records of every action an agent takes. By "transcript-grade" we mean a record sufficient to reconstruct, after the fact, what the model was asked, how it reasoned, and what it returned.
Agent runtimes today expose hooks inconsistently. Coverage varies between vendors and over time. There is no shared definition of what a complete hook surface looks like, what events must be emitted, or what metadata each event must carry. Without a standard, every enterprise rebuilds the same compliance plumbing per vendor, and every regulator has to evaluate every implementation from scratch.
AgentHook defines that standard.
SPEC.md: the wire-format specification (envelope, event types, metadata keys)GOVERNANCE.md: how the working group operates, perpetual licensing commitmentsCHARTER.md: formal stewardship terms, signedMEMBERS.md: founding members and invitee statusCONFORMANCE/: test rig (forthcoming)PROPOSALS/: change process (AHP, "AgentHook Proposal")
Pre-v1.0. Working group composition is being finalised. Public draft expected after founding members are confirmed in writing.