Multi-Agent System Patterns

A role is useful when it has a clear interface:

• what input it accepts
• what output it produces
• which decisions it can make alone
• when it must escalate

Without interfaces, agents overlap until they become interchangeable. That leads to redundant work, groupthink, and vague accountability.

Persona design still matters, but only after the interface is clear. Personality should shape tone and bias within a job, not replace the job definition.

Parallel generation is not enough. If every model sees the same context and gets the same objective, you often get the same answer with different phrasing.

For important work, use a three-step review pattern:

1. Write: one agent proposes the first answer
2. Cross-examine: another agent attacks weak assumptions
3. Judge: a coordinator decides what survives

This pattern is slower than a single pass, but it produces better judgment under uncertainty. Use it for strategy, deployments, or anything where shallow agreement is dangerous.

Teams need memory, but too much shared state creates confusion.

Store only what future work needs:

• durable decisions
• active work items
• current objectives
• known constraints
• review outcomes

Avoid dumping every draft into one giant memory store. Good shared state lowers ambiguity. Bad shared state becomes another hallucination surface.

A swarm looks impressive when many agents move at once. Production systems care more about sequencing than spectacle.

Useful orchestration answers:

• which roles can run in parallel
• which steps must block on review
• what happens on failure
• when a human must be pulled in

If you cannot describe the failure path, you do not have orchestration yet. You only have concurrency.