What are the main AI orchestration patterns?

Microsoft’s architecture guide names five patterns for coordinating multiple agents, chosen by matching the pattern to the shape of the coordination problem rather than picking a favorite.^[1]

• Sequential chains agents in a predefined, linear order. Each agent processes the previous one’s output, forming a pipeline of specialized transformations — a good fit when the steps themselves are fixed, like a document moving through drafting, review, and formatting.^[1]
• Concurrent runs several agents on the same task at once, each contributing an independent perspective. Useful when you want diverse analysis or a form of voting, not a division of labor.^[1]
• Group chat puts multiple agents into one shared conversation thread, with a chat manager deciding who responds next. A specific, more structured version of this is the maker-checker loop: one agent proposes, another evaluates against defined criteria, and the proposal cycles back with feedback until it is approved or the orchestration hits a maximum number of iterations.^[1]
• Handoff lets agents dynamically delegate: each agent can assess a task and either handle it or transfer it to a more appropriate specialist based on what it finds, the way a support ticket gets escalated to whoever actually knows the answer.^[1]
• Magentic is for open-ended problems with no predetermined plan. A manager agent builds and refines a task ledger as the work proceeds, developing goals and subgoals on the fly rather than following a route decided in advance — the pattern for “we don’t know the steps yet,” not just “we don’t know the answer yet.”^[1]

These describe how several agents split one job between them. A related but distinct taxonomy, from Anthropic’s guidance on structuring a single model’s calls, covers prompt chaining, routing, parallelization, orchestrator-workers, and the evaluator-optimizer loop — see workflows vs agents for that list.^[2] The two overlap in spirit: chaining resembles sequential, routing resembles handoff. But Microsoft’s set assumes multiple independent agents, where Anthropic’s assumes one model filling in different wiring.