Graphite and yellow processing nodes connected to a central routing hub
Capability orchestration is safest when the fan-out boundary and final verifier are chosen before the prompt leaves the local agent.

Choose Direct A2A Calls or Capability Orchestration

Hermes can call one explicit A2A peer with a2a_call, or use a2a_orchestrate to fan work out to peers advertising a requested capability. Direct calls optimize for control. Capability orchestration optimizes for breadth or routing convenience.

Official documentation: https://hermes-agent.nousresearch.com/docs/user-guide/messaging/a2a

Use direct calls when identity matters

Choose one peer when the workflow requires a specific credential owner, data boundary, specialist, cost profile, or service-level expectation. Discover the peer first, review its Agent Card, and send the smallest useful prompt.

Use capability orchestration when several approved peers can independently handle the same public-safe task. The documented modes are all, first, and best. The public A2A guide does not define ranking, ordering, quorum, consensus, or failover semantics for those labels, so do not build high-stakes policy on assumed behavior.

Fan-out can multiply disclosure, latency, and model cost. It also produces several self-reported answers that still need one local verification pass.

Orchestration checklist

  • Review every eligible peer before enabling fan-out.
  • Confirm the prompt may be shared with all selected peers.
  • Choose direct or capability-based routing explicitly.
  • Use synthetic or public data for the first trial.
  • Set a bounded time budget.
  • Define how contradictory results will be checked.
  • Keep one agent responsible for final verification.

Common pitfalls

  • Treating advertised capability as proven quality. Agent Cards are declarations.
  • Sending sensitive input through all. Every selected peer may receive it.
  • Calling fan-out consensus. Multiple outputs do not automatically become a reliable decision.
  • Ignoring duplicated cost. Each remote agent may make its own model and tool calls.

Verification

Run a harmless task through one explicit peer, then through the intended capability mode. Record which peers were contacted, compare results, and verify the answer against an independent source. Confirm the audit trail makes fan-out visible and that removing a peer from the approved set prevents it from receiving later work.