A good agent is a talented soloist: give it a goal, it plays until the piece is done. But some tasks are not solo pieces. Auditing a whole codebase, researching a market from twenty angles, migrating two hundred files: too big for one context, too varied for one specialist. That’s when you stop hiring a soloist and start assembling an orchestra.
That’s all agent orchestration is: making several specialized agents work together on one goal, with someone to keep them in time. The word sounds enterprise-y; the idea fits in one image, and we’ll keep it for the whole article. You’ll see: it’s not rocket science.
The conductor doesn’t play an instrument
Watch a conductor during a concert: they produce not a single note. They chose the program, decided who plays when, listen to everything, and fix the tempo when the brass rushes. Remove them and each musician still plays beautifully; together they drift apart.
An orchestrated system works the same way:
- The orchestrator is the conductor. Usually an agent itself, it takes the goal, splits it into parts, hands each part to the right specialist, watches the results come back, and assembles the final piece. It does almost none of the “real” work.
- The worker agents are the musicians. Each one is a complete agent: its own agentic loop, its own tools, and crucially its own context window. The violin doesn’t hear the sheet music of the oboe: each specialist works in its own bubble, focused, without drowning in the others’ details.
- The handoff is the relay between parts. When one agent passes work to another, it must pass the context along too: the goal, what’s been done, what remains. A botched handoff is the number one bug of multi-agent systems; the second violin can’t guess what the first was playing.
Together, this is a multi-agent system, and agent orchestration is the art of conducting it.
Five orchestras you’ll actually meet
Enough theory. Here are five real arrangements, each built on a pattern from our patterns collection.
1. The support desk. A cheap triage agent reads each incoming ticket and hands it to the billing specialist, the technical specialist, or a human. Each specialist has only its own tools and documentation. That’s routing: the conductor pointing to the section whose turn it is.
2. The newsroom. To produce a market study: three researcher agents fan out at the same time (one reads the competitors, one the regulations, one the customer reviews), a writer agent merges their notes into a draft, an editor agent critiques it against a checklist, and the draft loops back until it passes. That’s parallelization plus evaluator-optimizer: sections playing together, and the rehearsal where the conductor says “again, from bar 12”.
3. The code review. One finder agent hunts bugs, another checks performance, a third security. Then each finding goes to a verifier agent whose only job is to try to refute it: only confirmed findings reach you. Coding tools work exactly like this under the hood (Claude Code’s subagents and workflows, for instance). Skeptical second opinions are much cheaper when they’re agents.
4. The migration. “Convert these 214 files to the new API.” The orchestrator lists the files, spawns a worker per batch, each worker converts its files in its own context, the orchestrator collects and runs the tests. That’s orchestrator-workers: the score is trivial, the sheer volume is the problem, and twenty musicians play twenty stands at once.
5. The deep research. Anthropic described how its Research feature works: a lead agent reads your question, spawns several searcher agents that explore different angles in parallel, then synthesizes their findings with citations. Their measured result: the multi-agent version beat their best single agent by 90% on internal research evals. The same write-up gives the price tag, and we’ll get to it, because it’s the honest part.
You’ll find the same arrangements in every orchestration toolkit, whatever the brand: LangGraph, CrewAI, the Microsoft Agent Framework, the OpenAI Agents SDK with its handoffs, or Claude Code’s subagents. The vocabulary changes; the orchestra doesn’t.
The word of honesty: most gigs need a soloist
Before you hire an orchestra, three warnings, from people who run them in production:
- It’s expensive. Every agent carries its own context, its own tool calls, its own retries. Anthropic’s own numbers: an agent burns roughly 4× the tokens of a chat, and a multi-agent system roughly 15×. The orchestra must create enough value to pay fifteen musicians.
- It’s harder to debug. A wrong answer might come from the researcher, the writer, the editor, or a handoff between them. Without per-agent traces and evals, you’re conducting blind.
- It only shines on decomposable work. Parallel musicians help when the piece splits into independent parts (read twenty sources, convert two hundred files). A deeply sequential task, where each step depends on the previous one, gains nothing from more players; sometimes agents even step on each other, like two pianists at one keyboard.
The rule of thumb is boringly simple: start with one agent. Move to an orchestra when a single context window genuinely can’t hold the task, or when independent skeptical opinions are the point (reviews, verifications). Not because the demo looked impressive.
The recap table
| The orchestra | The system | The pattern |
|---|---|---|
| The conductor | Orchestrator agent: decomposes, dispatches, assembles | orchestrator-workers |
| Pointing to a section | Sending the task to the right specialist | routing |
| Sections playing together | Independent agents working in parallel | parallelization |
| “Again, from bar 12” | A critic agent loops the work until it passes | evaluator-optimizer |
| Passing the melody | Handoff of task + context between agents | handoff |
| The full concert | Several complete agents, one goal | multi-agent |
In a nutshell
Agent orchestration is the conductor’s job applied to AI: one agent that decomposes, dispatches to specialists who each work in their own context, and assembles the result. It’s the right tool for work that is too big or too parallel for one agent, and a 15× token bill for work that isn’t.
One question remains, and it’s a big one: who writes the score? Today our conductor improvised, deciding live who plays next. Tomorrow we’ll see what happens when you write the score down in advance, notes, repeats and all. Spoiler: it involves graphs, and it’s still not rocket science.
A word you don’t know? The glossary defines every AI term, in plain language.
