OpenAI Agents SDK Integration¶

Verify a manifest before an agent handoff or protected tool operation. Start with the runnable verification gate; it uses synthetic inputs and needs no model API key.

Connect a handoff¶

Keep the signed manifest, approved verification context, and revocation state in application-owned context. A model-generated manifest ID is only a locator and must not choose the recipient's trust anchors or expected runtime hashes.

The Agents SDK exposes handoff(..., on_handoff=...). Perform checks at the start of on_handoff, before application side effects, and raise on rejection: a successful callback return allows the transfer to continue. is_enabled can control availability but cannot authorize handoff arguments that have not yet been generated. See the SDK's handoff documentation.

Authenticate the signed object against approved issuer keys and current runtime observations. A handoff to another agent also needs the application's delegation and scope checks. A valid manifest does not independently authorize the requested transfer or input data.

Test the boundary¶

Test accepted, unknown-key, revoked, and changed-artifact cases before running a live model. Confirm each rejection prevents the receiving agent and downstream tools from executing. Cover direct tool calls separately; a handoff hook is not a gate for every tool path.

Pin the Agents SDK version used by your application and verify its hook behavior. This page supplies the verification pattern; the framework does not automatically apply Agent Manifest checks.