Choose a coordination pattern that makes authority and failure visible

Supervisor and specialist pattern

A supervisor assigns work to specialists and assembles their results. It provides a clear control point and is often the easiest topology to explain. The supervisor can become a bottleneck, single point of failure, and source of biased aggregation.

• Best when tasks are heterogeneous and centrally decomposable
• Require evidence-aware routing and explicit completion criteria
• Test whether the supervisor can detect specialist failure
• Keep consequential approval outside the supervisor model

Pipeline and handoff pattern

A pipeline passes an artifact through roles such as researcher, analyst, drafter, verifier, and formatter. It works well for stable stages and bounded transformations. It can amplify early assumptions and discard source context as artifacts are summarized.

Retain original evidence and allow later roles to reopen prior stages rather than forcing blind acceptance of the handoff.

Blackboard pattern

Agents coordinate through shared state. This supports asynchronous work and opportunistic contributions. It requires write ownership, versioning, conflict detection, provenance, access control, and protection against poisoned memory.

Peer deliberation and critique

Peers propose, critique, revise, or vote. The pattern can expose uncertainty when participants are meaningfully independent. It can also produce groupthink, persuasion contests, duplicated cost, and endless convergence cycles.

Define the decision rule, maximum rounds, minority-evidence treatment, and human escalation before running the debate.

Auction and contract-net patterns

Participants bid for tasks based on capability, cost, confidence, or load. This is useful in dynamic allocation but introduces incentive design. A bidder may optimize the score rather than the actual outcome. The coordinator needs admission controls, budget limits, and post-task quality feedback.

Swarm-like local coordination

Participants follow local rules and coordinate through shared signals or neighborhood state. This can increase adaptability and reduce central bottlenecks. It also makes global causality and convergence harder to inspect. Use it only when decentralized coordination is intrinsic to the problem.

Hybrid patterns

Real systems often combine structures: a supervisor authorizes a bounded swarm search; specialists write to a blackboard; a deterministic workflow pauses for human approval before execution. Name the governing mechanism at each layer so authority does not become ambiguous.

Compare against a strong baseline

Research on multi-agent failures and team underperformance shows why topology must be evaluated rather than assumed. [S4] [S5] Compare quality, cost, latency, reviewer effort, security behavior, and recovery against one strong agent or deterministic workflow.

Research boundary

Patterns are not performance guarantees. Models, prompts, data, tools, and task structure can change outcomes. Use representative cases and retain human ownership of consequential decisions.