loading…
Distribute work across specialized decision-making actors only when the boundaries are real.
A multi-agent system contains multiple actors with their own role, context, state, or authority. It is justified when specialization, independent ownership, isolation, or cross-organization interoperability matters—not merely because the task has several steps.
| Topology | How it works | Risk |
|---|---|---|
| Supervisor | One orchestrator delegates to specialists | Supervisor bottleneck |
| Sequential handoff | Each agent passes an artifact to the next | Context loss at boundaries |
| Peer collaboration | Agents coordinate without one permanent controller | Harder convergence and debugging |
| Critic / reviewer | One agent evaluates another's artifact | Correlated errors or performative debate |
A multi-agent system can communicate through in-process calls, queues, HTTP APIs, or a standardized agent protocol. A2A is designed for communication and interoperability between independent, potentially opaque agents built with different frameworks or operated by different organizations.
Multi-agent = how the system is organized A2A = one option for how independent agents discover and communicate
An Agent Card advertises capabilities and interaction details. Task-oriented exchanges can carry status updates and artifacts across long-running work. These features are most valuable at organizational or framework boundaries; they are often unnecessary for three small agents inside one application process.
Decide whether agents share one state store or receive isolated task packets. Handoffs should include the goal, evidence, completed work, unresolved questions, constraints, and expected artifact—not a vague "continue from here."
More agents create more prompts, context copies, network calls, permissions, traces, and opportunities for disagreement. Measure whether specialization improves end-to-end success enough to justify that tax.