Multi-agent handoff benchmark

Source

This synthetic case starts from one brief and a small source.html fixture. The named producer and consumer are roles in the regression case, not claims about external agent services or customer usage.

Change

The planning role produces a versioned start payload with an embedded design contract and a canonical static quality report from the supplied HTML. Both payloads are serialized and parsed before the implementation-review role receives and validates them. The benchmark checks that this transfer changes neither contract nor permission state.

Verification

Run:

design-ai benchmark multi-agent-handoff --strict

The case validates the planning role's start contract and quality report, then the implementation-review role's received copies. It compares the transferred objects for exact semantic equality, checks the approval gate, and confirms that a clean static source still carries the runtime unverified finding.

Permission boundary

The roles exchange in-memory JSON only. No MCP server, model provider, repository, browser, local evidence directory, or external service is contacted. Editing and publishing remain explicit downstream approvals in the transferred contract.

Remaining risk

The case proves portable contract integrity inside the local runner. It does not prove delivery guarantees of a real agent transport, implementation correctness, or runtime quality. A production handoff still needs transport evidence, approved implementation, and post-change browser verification.

Claim boundary

This synthetic role transfer does not prove external agent adoption or transport reliability. It proves only the local envelope digest, consumer validation result, and preserved approval boundary for the packaged payloads.