RC-401b · Module 2

Communication Protocols

3 min read

In a well-designed agent team, all information flows through the lead. Specialist A does not pass output to Specialist B directly. The lead reviews A's output, extracts what B needs, and includes it in B's prompt. This hub-and-spoke communication model costs more tokens than direct agent-to-agent messaging, but it provides two properties you cannot get any other way: complete visibility and error isolation.

Complete visibility means the lead sees every piece of data moving through the system. When something goes wrong — and at 0.95^N reliability per agent, something will go wrong — the lead knows exactly which agent produced the bad output and which downstream agents consumed it. Error isolation means a failure in Agent A does not propagate silently through Agents B, C, and D before anyone notices. The lead catches it at the routing layer.

Do This

  • Route all inter-agent communication through the lead — visibility is worth the token cost
  • Define explicit handoff schemas: what data each agent produces and what the next agent expects
  • Include context summaries in handoff prompts — the receiving agent does not share the sender's memory
  • Set convergence criteria for debate rounds: "agree on top 5 findings" or "maximum 3 rounds"

Avoid This

  • Let agents communicate directly — you lose traceability and error isolation
  • Pass raw agent output to the next agent without lead review and extraction
  • Assume agents share context — each maintains its own independent context window
  • Run debate rounds without explicit stop conditions — agents will iterate until your budget is gone