Orchestrator (Agent Pattern)

An orchestrator is a design pattern in multi-agent AI systems where one agent acts as a manager — planning the work, delegating tasks to specialist agents, reviewing their output, and coordinating the overall effort. The orchestrator never does the work itself. It manages the process.

How the orchestrator pattern works

1. Receive goal: The orchestrator receives a high-level objective from the user
2. Plan: It decomposes the goal into specific sub-tasks
3. Delegate: It assigns each sub-task to a specialist agent best suited for it
4. Monitor: It tracks progress and handles any failures or unexpected results
5. Review: It evaluates the quality of each sub-agent's output
6. Synthesise: It combines all results into a coherent final deliverable
7. Deliver: It presents the result to the user

Why not just use one agent?

A single agent doing everything faces three problems. First, context window limits — a complex task can fill the window before the agent finishes. Second, quality degradation — an agent that researches, writes, edits, and formats all in one go tends to lose focus. Third, speed — one agent doing tasks sequentially is slower than multiple agents working in parallel.

The orchestrator solves all three by delegating to specialists, each with their own fresh context window and a narrow, focused objective.

Real-world example: AI content pipeline

An orchestrator receives the brief: "Write a blog post about AI in healthcare."

It delegates to: - Research agent: Find 5 recent statistics and trends in AI healthcare adoption - Writer agent: Write an 800-word article using the research brief - Editor agent: Review for accuracy, tone, and brand voice compliance - Formatter agent: Convert to the required format with metadata

The orchestrator reviews each handoff, requests revisions if quality is insufficient, and delivers the final article.

Key consideration: The orchestrator itself uses tokens for planning and coordination. For simple tasks, the overhead of an orchestrator exceeds the benefit. Use this pattern when tasks have 3+ independent parts that benefit from specialisation.