Agent-to-agent (A2A)
Agent-to-agent (A2A) is the emerging open standard for agents that transact with each other directly. Switchboard's posture: REST and MCP are the shipped agent surfaces today; A2A ships as an adapter on the same state layer as the protocols mature. Same orders, same approval gates, same audit trail; a different wire format, not a different market.
The agent card
Each tenant scope publishes an agent card at a well-known path, so a counterparty agent can discover what it can do here before it says a word.
GET https://api.switchboard.simulmedia.com/.well-known/agent-card.json
- The card advertises the marketplace skills available to that scope (approvals on the seller side; catalog search, avails, order create and track) and the auth it expects.
- Scopes mirror the REST API: what your credentials can reach over REST is exactly what an agent can reach over A2A. See Authentication.
Task lifecycle mapping
A2A models work as tasks. Tasks map onto the canonical order lifecycle; no state exists on one side that cannot be expressed on the other.
| Order state | A2A task state |
|---|---|
submitted / validating | submitted / working |
needs_confirmation, negotiating | input-required |
seller_review | working |
placed / completed (including partial) | completed |
rejected / canceled / failed | rejected / canceled / failed |
The division of labor is protocol-independent: deterministic rules clear the routine path, agents absorb the exceptions inside bounds you set, and people own the judgment calls. An agent negotiating over A2A still cannot skip
seller_review.Transport and auth
- JSON-RPC 2.0 over HTTPS; streaming updates over SSE, plus push notifications riding the same notifier as webhooks.
- OAuth is unchanged: the same client-credentials flow and scopes as REST and MCP. No separate identity system for agents.
- Status: designed, on the committed roadmap, and sequenced after the v1 REST surface. Building an A2A-speaking agent and want early access? Talk to us.