Multi-Agent Architecture Patterns
Once you've decided to use several agents, the next question is: how should these agents communicate with each other? The four common patterns below give different answers to that question.
The Supervisor Pattern
One central agent (the "supervisor") receives the user's request, decides which specialized agent should handle it, delegates the work, and consolidates the result. Worker agents don't talk to each other directly, only to the supervisor.
The Hierarchical Pattern
An expanded version of Supervisor: several layers of supervisors stack on top of each other — a senior supervisor splits work among several "mid-level supervisors," and each of them manages their own team of worker agents. Suited to large organizations with multiple work domains.
The Peer-to-Peer Pattern (Decentralized)
There is no central supervisor; agents communicate directly with each other (often through a protocol like A2A) and decide jointly. More flexible, but coordination and predicting the system's behavior become harder.
The Pipeline Pattern (Linear)
Agents work in a fixed order, one after another — each one's output becomes the next one's input (for example: an information-extraction agent → a validation agent → a transaction-execution agent). The simplest pattern to implement and debug, but less flexible for non-linear paths.
The Debate / Consensus Pattern
Several instances of one agent (or several agents with different viewpoints/prompts) independently answer the same question, then critique each other's answers and refine their positions over one or more rounds of debate. In the end, a final answer is chosen through voting or a judge agent's summary. Unlike Supervisor, where one decision-making authority is defined from the start, "correctness" here emerges from the confrontation of several viewpoints.
This pattern has a high compute cost (several times more model calls than a simple answer), but for sensitive decisions with a high error risk (like final approval of a large financial transaction, or assessing content quality), it can meaningfully improve accuracy, since one agent's error is likely to be corrected by the others.
Comparison & Choosing One
| Pattern | Complexity | Best suited for |
|---|---|---|
| Supervisor | Medium | Most practical scenarios; a good default starting point |
| Hierarchical | High | Large organizations with several distinct work domains |
| Peer-to-Peer | High | Collaboration between independent agents from different organizations |
| Pipeline | Low | Sequential, predictable processes |
| Debate / Consensus | High (and costly) | Sensitive decisions with a high error risk |
FAQ
Which pattern is best to start with?
For most teams, Supervisor is a good starting point: structured enough to be debuggable, and flexible enough for most scenarios.
Can patterns be combined?
Yes, this is common — for example, a Supervisor that is itself one node of a larger Pipeline.