Agent-to-agent trust

Two AI agents, each a Verifiable Service, want to work together. Neither trusts a name or an API key. The sequence:

1. Find: agent A searches the graph for a counterpart by what it proves, for instance an issuer accredited for a given schema in a given ecosystem (verana.graph.search.did with the ecosystem and schema facets), and picks a DID.
2. Resolve: verana.idx.vt.resolve on that DID. Trusted or nothing happens.
3. Connect: A's VS Agent opens a DIDComm connection to B's; B's agent resolves A in turn and refuses if untrusted. Both backends see didcomm.connections.state-updated.
4. Negotiate access: over DIDComm, B issues A a token, or asks A's agent for a credential A holds (its accreditation as a client of B's ecosystem, a badge of its operator).
5. Call: A calls B's MCP or A2A endpoint, declared in B's DID Document, with that material.
6. Keep checking: verdicts expire; the agents re-resolve each other before new sessions and stop on an untrusted verdict.

The mcpi-two-vs-agents example runs this end to end with two agents on testnet.

Next

• Delegated execution and limits.
• Understand: Verifiable AI agents.