Multi-agent coordination

Multi-agent coordination divides work among agents and defines how they share context. A supervisor applies a coordination pattern and decides when the combined work is complete.

Supervision and context

Multi-agent coordination divides a task among agents and decides how their results contribute to a common outcome. Use it when separate contexts, specialized capabilities, or parallel work provide enough benefit to justify the coordination cost.

Choosing a supervisor

A Workflow coordinates agents through steps defined in code. It suits a known sequence, explicit recovery policies, and compensation of earlier business actions.

An Autonomous Agent coordinates through capabilities declared in its definition. The model decides which worker or peer to use, while the runtime manages tasks and execution. It suits work whose next step depends on intermediate findings.

Keep worker agents focused so they can be reused in different processes. Place decisions about sequence, retries, and completion with the supervisor instead of distributing that policy across unrelated workers.

Context and work boundaries

A coordinator can delegate isolated subtasks and combine their typed results, transfer ownership through handoff, lead a team exchanging messages, or moderate a conversation. These choices share different context and have different completion rules.

Use isolation when independent reasoning matters. Share information when agents need to respond to each other’s work. A shared session or team does not imply that every participant sees every other participant’s private history.

A person can also be part of the process. A workflow pauses while waiting for an approval and resumes when a command supplies it. Validate the waiting state and the caller’s authority before accepting the decision.

Limits and trade-offs

Start with one agent where it is sufficient. More agents add model calls, latency, context management, and failure paths. A model-driven supervisor is less predictable than a fixed sequence, so bound its iterations and evaluate the complete process.

Durable coordination preserves progress; it does not make external actions exactly once. Retried tools and workflow calls still need safe duplicate handling.

Coordination patterns

Coordination patterns describe how agents divide work and exchange context. Choose the pattern from the task’s dependencies and information boundaries, rather than adding agents solely to mirror organizational roles.

Four patterns

Pattern How work moves Main trade-off

Sequential

One stage passes its output to the next. Implement a fixed sequence with workflow steps, or transfer task ownership through handoff.

Simple dependencies, but early mistakes influence later stages and work does not run in parallel.

Delegative

A coordinator assigns isolated subtasks to workers and combines their results.

Parallel work and independent reasoning, with decomposition and synthesis concentrated in the coordinator.

Collaborative

Agents exchange information while working. Teams share tasks and messages; moderation structures turns and completion.

Useful for interdependent work and review, with coordination overhead and a risk of early convergence on one answer.

Emergent

Many agents contribute to shared state, and a selection process turns contributions into a result.

Broad exploration, with less predictable outcomes and a need for explicit selection criteria.

A shared session is a context-sharing mechanism, not a pattern by itself. It can support several patterns, depending on who writes, reads, and controls the next action.

Combining patterns

A sequential process can delegate one stage to parallel workers. A coordinator can delegate to a team whose members collaborate, or a debate can pause for independent research. Keep each boundary explicit: what work is assigned, what context crosses it, what result returns, and who decides completion.

The pattern adds no delivery or recovery guarantee of its own. Those come from its workflow or Autonomous Agent supervisor. External side effects still need safe retry behavior.

Next steps