A controlling agent that directs sub-agents, manages state, and decides when a task is complete.
Orchestration is the meta-pattern of multi-agent systems. The orchestrator doesn't do the work — it coordinates the agents that do. It maintains the overall task state, decides what to do next based on what sub-agents have produced, handles failures, and determines when the task is complete.
The orchestrator maintains a persistent representation of the overall workflow state. It dispatches tasks to sub-agents, waits for results, evaluates whether results meet the criteria to proceed, and decides the next step. This decision loop continues until the orchestrator determines the goal is achieved, a termination condition is met, or a failure requires escalation. The orchestrator must be resilient to sub-agent failures and must not lose workflow state if it is restarted. All state should be persisted to durable storage at each decision point. The orchestrator's scope of authority must be explicitly bounded: what actions can it authorise sub-agents to take without human approval?
A legal firm uses an orchestrator to manage contract review workflows. When a contract arrives, the orchestrator dispatches it to specialist agents: a clause extraction agent, a risk identification agent, a compliance check agent, and a precedent comparison agent. As results arrive, the orchestrator maintains a running risk register. If risk identification returns high-severity findings, the orchestrator pauses the workflow and routes a summary to the reviewing partner before proceeding. Once all agents complete, the orchestrator compiles the final report and routes it to the file management system. Every decision point is logged with the orchestrator's reasoning.
Without orchestration, multi-agent systems have no coherent state, no failure recovery, and no consistent definition of 'done'. Tasks start but don't complete. Failures go undetected. Outputs from different agents contradict each other without anyone noticing. Orchestration is the pattern that turns a collection of agents into a system.
How this pattern fails in practice — and what to watch for.
The orchestrator loses track of workflow state when it crashes or is restarted. It either restarts the entire workflow (causing duplicated actions) or cannot recover at all. Sub-agents that completed successfully have their outputs discarded.
A sub-agent returns a result that indicates failure in the content of its output but the orchestrator interprets it as success because the return format was valid. The orchestrator proceeds through the workflow, building on a failed foundation.
The orchestrator is given the ability to dynamically replan the workflow. When original sub-agents fail, it creates new tasks that fall outside the original workflow's authorised scope. Actions are taken that were never approved.
Seven things to verify before deploying this pattern in production.
Orchestration is covered in depth across AIDA (D4 and D6), CAIG (accountability and scope governance), and AIMA. The CAIAUD exam specifically tests auditors on orchestration state management and scope governance — can you trace a completed workflow back through orchestration logs and verify that no unauthorised actions were taken?
The AIDA certification covers all 21 agentic design patterns with a focus on deployment safety, governance, and the PSF. Free to attempt.