Orchestration vs choreography for AI agents: who owns the plan
Updated November 2026
Orchestration puts one component, a supervisor agent or plain code, in charge of the workflow. That component decides which agent runs next and collects the results. Choreography removes it, so each agent reacts to events on topics it subscribes to and publishes events of its own. Choose orchestration when one place should own the plan and its status. Choose choreography when agents change often and a central node would become a bottleneck or a single point of failure.
At a glance
| Orchestration | choreography | |
|---|---|---|
| Where the workflow logic lives | Orchestration keeps the plan in one component, and in Microsoft's description each service isn't aware of the overall workflow. | Choreography decentralizes workflow logic and distributes responsibilities across the components of a system. |
| How agents are coordinated | Orchestration can be driven by a model or by code. The OpenAI Agents SDK describes letting the LLM decide what steps to take as one option and determining the flow via code as the other. | Choreography routes work through broker topics. In Solace Agent Mesh, a topic keyed by an agent name is a request channel routed to that agent, and a topic keyed by a task identifier is a response channel scoped to one in-flight call. |
| State and memory | Orchestration lets the coordinator consolidate the status of a transaction from the results of individual operations. For long-running work, Microsoft advises you to persist shared state externally instead of relying only on in-memory context. | Choreography can put shared state on a topic. Confluent describes a blackboard that becomes a data streaming topic of messages that worker agents produce and consume. |
| Typical failure mode | Orchestration concentrates risk in the coordinator, which under load can be the single point of failure. | Choreography can produce feedback loops or event storms when many services react to each other's events. |
| Tracing | Orchestration gives you one process to wrap in a trace. The OpenAI Agents SDK lets you put multiple calls into one trace by wrapping the code in a trace() block. | Choreography spreads one request across asynchronous steps. OpenTelemetry span links let you associate one span with others, implying a causal relationship. |
| Protocols | Orchestration is the shape of the A2A overview's own example, where a trip-planning request involves orchestrating multiple specialized agents. | Choreography over a broker can still use A2A. Solace Agent Mesh keeps the full list of every topic, publisher, and subscriber in its A2A protocol reference. |
How does each one coordinate agents?
In orchestration, a coordinator receives the request and hands out the work. In the distributed-systems definition, incoming requests flow through the orchestrator as it delegates operations to the respective services. For agents, the coordinator is a supervisor or manager agent, or plain code. In CrewAI's hierarchical process, tasks are delegated through a structured chain of command, and a manager must be specified as either a manager LLM or a custom manager agent. Microsoft notes that some patterns route deterministically while others let agents choose their own routes.
In choreography, no component holds the plan. The pattern asks you to have each service decide when and how to process an operation instead of depending on a central orchestrator. Microsoft's choreography pattern page shows an implementation that uses a message broker to coordinate communication between services. When an operation succeeds, the service can publish a message so another service continues the workflow, and when it fails, the service publishes a failure message. Confluent shows that the orchestrator-worker pattern can be adapted so the agents become event-driven, with workers that inherit a Kafka consumer group's functionality for coordination, scaling and fault recovery. The event-driven pattern page covers this design in detail.
When should you choose orchestration?
- The steps must run in a fixed order. In Microsoft's sequential pattern, the next agent is defined as part of the workflow and isn't a choice given to agents.
- You want predictable speed and cost. The OpenAI Agents SDK says orchestrating via code makes tasks more deterministic and predictable in speed, cost and performance.
- A person should approve the plan before work starts. Microsoft Agent Framework's Magentic orchestration lets users review and approve the proposed plan before execution.
- Each coordination round should leave a record. Magentic emits a progress ledger event once per coordination round, carrying whether the team is in a loop, the next speaker and the instruction sent to them.
When should you choose choreography?
- Components are replaced often. Microsoft lists the case where you frequently update and replace components as a reason to use the pattern.
- Agents join and leave groups. In Confluent's event-driven hierarchy, agents can be added or removed from sibling groups without the individual agents having to manage the change.
- One agent needs several replicas. In Solace Agent Mesh, replicas of a component bind to the queue by the same name, and the event broker delivers each message to exactly one replica.
- Teams or domains must stay loosely coupled. Microsoft points to communication between bounded contexts that requires loose coupling across domains.
Can you use them together?
You can, and a common split follows domain boundaries. Microsoft suggests choreography between bounded contexts and, for work inside a single bounded context, an orchestrator pattern depending on complexity and team preference. The saga pattern treats choreography and orchestration as two ways to coordinate the same kind of transaction. Solace Agent Mesh puts an orchestrating agent on top of an event broker. Its Orchestrator is a regular agent configured to delegate, and replicas of one agent share the queue that carries requests to that agent.
To trace work across the boundary, pass one trace identifier through each hop. W3C Trace Context starts from the point that a distributed trace traverses more than one component and must be uniquely identifiable across all participating systems. The guide on debugging multi-agent systems covers the practice.
What should you watch out for?
- A synchronous coordinator can block. Anthropic says its lead agents execute subagents synchronously, so the lead agent can't steer subagents and the whole system can wait on one subagent.
- Events can arrive twice or out of order. Microsoft warns that at-least-once delivery and retries can produce duplicates and advises idempotent consumers that track stable message identifiers.
- Delivery mode matters. With Solace direct subscriptions, if no replica is connected when a message is published, the message is dropped.
- Debugging gets harder as participants grow. Microsoft's saga guidance says debugging sagas can be complex as the number of participating services grows.
What else should you know?
Is a supervisor agent orchestration or choreography?
A supervisor is orchestration, because one agent decides who works next. See supervisor agents for the agent version of a central coordinator.
Do I need multiple agents at all?
Not always. Microsoft advises that if a single agent can reliably solve your scenario, you should consider that approach.
Where do A2A and MCP fit?
The A2A overview says A2A crosses the boundary between agents, while MCP stays within one agent and connects it to APIs and enterprise applications.
How do I follow one request through event-driven agents?
Give the request a trace identifier and carry it in every message. In OpenTelemetry, span context holds the Trace ID, and it is the part of a span that is serialized and propagated alongside distributed context and baggage.
Which sources does this page use?
- Choreography Pattern (learn.microsoft.com), reviewed November 2026
- Agent orchestration (openai.github.io), reviewed November 2026
- The Event-Driven Mesh --> Support (docs.solace.com), reviewed November 2026
- AI Agent Orchestration Patterns (learn.microsoft.com), reviewed November 2026
- Four Design Patterns for Event-Driven, Multi-Agent Systems Data architecture tha (confluent.io), reviewed November 2026
- Tracing (openai.github.io), reviewed November 2026
- Traces (opentelemetry.io), reviewed November 2026
- What is A2A? (a2a-protocol.org), reviewed November 2026
- Processes (docs.crewai.com), reviewed November 2026
- Microsoft Agent Framework Workflows Orchestrations (learn.microsoft.com), reviewed November 2026
- Pattern: Saga Microservice Architecture (microservices.io), reviewed November 2026
- Step 3: Build Multi-Agent Orchestration --> Support (docs.solace.com), reviewed November 2026
- Trace Context Trace Context. W3C Recommendation 23 November 2021 (w3.org), reviewed November 2026
- How we built our multi-agent research system \ Anthropic (anthropic.com), reviewed November 2026
- Saga Design Pattern (learn.microsoft.com), reviewed November 2026