Why architecture matters here
Multi-agent systems fail in three ways: conflict (two agents fight for the same task), silence (nobody picks up the work), and cost blowup (agents talk without acting). Each requires an architectural piece to prevent — arbitration, coordinator assignment, and budgets.
The architecture matters because you cannot rely on the agents themselves to solve these. Even great planners misbehave when there is another agent in the loop. The orchestration plane must impose structure.
With that structure, specialization delivers real leverage; without it, multi-agent is worse than one big agent.
The architecture: every piece explained
The top strip is the runtime. Coordinator assigns tasks to agents based on capabilities and current load. Message bus exchanges typed envelopes so agents communicate in structured, auditable ways rather than free-form chat. Agent A / B / C are role-specialized (retriever, analyst, writer, verifier). Capability registry holds each agent's declared skills and their policy scope.
The middle row is control. Arbitration handles conflicts — two agents proposing incompatible actions — either by rule (precedence) or by a judge agent. Shared memory holds decisions, facts, and open issues in a structured store. Escalation sends stuck decisions to a human. Guardrails + policy enforce cross-agent limits so no single agent can bypass safety.
The lower rows are ops. Trace + audit records every message and decision for review. Cost + budget caps tokens and time per task so agent debates cannot burn unbounded compute. Ops includes deploy versioning, failure containment (one bad agent version does not corrupt the system), and replay from trace for debugging.
End-to-end flow
End-to-end: a product-analytics task arrives. Coordinator assigns retrieval to the Retriever agent, analysis to the Analyst, and drafting to the Writer. Retriever posts findings on the message bus; Analyst consumes and posts conclusions; Writer drafts a report. A Verifier catches a questionable statistic; arbitration rule sends it back to Analyst for confirmation. Shared memory has the decision trail. Cost budget is 85% used; coordinator ends the loop. Audit trace confirms compliance; if a stakeholder asks how a claim was made, replay shows the exact reasoning. The report ships.