Meet-Me

Pattern 1 of 2. Public, manual agent-to-agent meeting coordination. Pattern 2 (authenticated, sandboxed) is in The-Lobby.

Mechanism

1. Principal publishes a skill manifest at a stable URL — e.g. https://<domain>/agents.md (and/or /.well-known/agents.md, plus a JSON twin at /agents.json).
2. Email signatures, business cards, and profile pages link to it.
3. A receiving agent fetches the manifest, parses the relevant skill, and acts on it (calendar link, MCP endpoint, routing rule, etc.).

Manifest shape

Hub-and-spoke. A top-level agents.md enumerates skills; each skill is a sibling file:

<domain>/
├── agents.md                          # hub: principal identity + skill index
└── agents/
    ├── meeting-coordination.md
    ├── press-inquiry.md
    ├── intro-request.md
    └── support-routing.md

Each skill file specifies: required inputs, available calendar/MCP endpoints, scopes, working hours, hard rules, fallbacks. See agents.md and agents.json for the worked example.

Dual-track signature

Two equivalent entry points in the signature: one calendar URL for humans, one manifest URL for agents.

V1 workflow

Agent reads inbound email → detects agent link in signature → fetches manifest → reads skill definition (availability data, booking link, meeting preferences) → executes booking action → reports completion to principal.

Threat model

Public manifest. No identity on the requester. No audit trail. Concretely:

• Reconnaissance leakage — anyone can scrape working hours, travel windows, delegated inboxes.
• No requester authentication — protection is whatever already sits behind the calendar link.
• No transcript — disputed bookings have no signed record of which agent acted.

Mitigations (without leaving Pattern 1)

• WAF / bot management on the manifest endpoint. Per-IP rate limits, headless-browser challenges, ASN deny-listing, optional reduced-detail variant for low-trust requesters.
• Human-in-the-loop confirmation on writes. Manifest advertises availability freely; the booking action holds as pending until the principal confirms (push, email, tap-to-approve). Caps blast radius regardless of who scraped the manifest.

When stronger guarantees are needed (verified requester identity, scoped one-shot capability, signed transcript) → Pattern 2 / The-Lobby.

Files

File
index.html	Email-signature wireframe
agents.md	Worked skill manifest (human-readable)
agents.json	Same content, structured
diagrams/	Diagram assets

Status

Notes / sketch. Not a spec. Prior-art pointers welcome.

License

MIT.