What is the orchestrator-worker pattern?
The orchestrator-worker pattern is a multi-agent design in which a lead model breaks a task into subtasks at run time, hands each to a worker model with its own context, and combines the workers' results into one answer.
Anthropic's "Building effective agents" (December 2024) named orchestrator-workers as one of five common patterns. What sets it apart from simple parallel fan-out is that the subtasks are not fixed in advance: the orchestrator decides them from the input, which suits jobs such as code changes spread across many files or research that pulls from many sources.
OpenAI's practical guide to building agents describes a close relative, the manager pattern, in which a central agent calls specialist agents as tools and keeps control of the conversation.
What the evidence says
Anthropic's write-up of its multi-agent research system (June 2025) reported that a lead agent on Claude Opus 4 with Claude Sonnet 4 workers beat a single Opus 4 agent by 90.2% on an internal research eval. The same post put the cost plainly: agents use about 4 times the tokens of a chat, and multi-agent systems about 15 times.
Google Research's study of 180 agent configurations (January 2026) found that coordination helps a lot on parallel tasks but cut performance by 39% to 70% on sequential reasoning. Independent agents with no cross-checking amplified errors about 17 times, while centralised coordination contained errors more effectively than the other set-ups tested, which favours an orchestrator over a group of peers when a team does split the work.
When it fits
Breadth-first work
Research, audits and comparisons where many independent threads can run at once.
More than one context window
Tasks whose material will not fit in one window, where each worker reads its share and returns a summary.
Many complex tools
Workers can each hold a narrower tool set, which keeps tool selection accurate.
Poor fit
Tightly coupled work such as most coding, where Anthropic found few truly parallel subtasks, and any task that is mostly sequential.
The 2026 correction
Anthropic's October 2026 guidance on folding subagents back into the main loop argues that, for many enterprise agents, one loop with progressively disclosed skills and tools beats orchestrator-plus-subagent designs on quality, cost and latency. It keeps subagents for context isolation, parallelism and conflicting instructions, and warns against subagents shaped like an org chart. The broader single versus multi-agent trade-off has its own comparison page.
How Motiontic's engineers apply it
Motiontic's engineers start with a single agent and add an orchestrator only when an eval shows the single loop running out of context or time on genuinely parallel work. Workers get scoped instructions, a narrow tool set and a required output format, and the orchestrator's merge step is evaluated on its own.
Orchestrator-worker compared with parallel fan-out
| Parallel fan-out | Orchestrator-worker | |
|---|---|---|
| Who defines the subtasks | The developer, in code | The orchestrator model, per input |
| Number of workers | Fixed | Varies with the task |
| Fits | Known checks run side by side, such as several review lenses | Open-ended research or changes across many files |
| Token cost | Predictable | High and variable |
| Main risk | Missing a subtask nobody planned for | Poor task splits and a weak merge step |
Common questions
What is the difference between orchestrator-worker and subagents?
Subagents are the workers. The orchestrator-worker pattern describes the whole arrangement: a lead agent that plans, spawns subagents with their own context, and merges what they return.
Is the orchestrator-worker pattern expensive?
Usually. Anthropic measured multi-agent research runs at about 15 times the tokens of a chat, so the pattern pays off only when the task is valuable and highly parallel.
How is orchestrator-worker different from a handoff?
In orchestrator-worker, the lead keeps control and calls workers for parts of the job. In a handoff, control of the conversation moves to another agent, which carries on from there.
Related
Single Agent vs Multi-Agent
One loop with good tools, compared with several agents splitting the work.
Subagent
A helper agent with a clean context that hands back a short answer.
Agent handoff
One agent passes the conversation to another, which carries on from there.
Agent teams
Several full agent sessions on one project, claiming tasks and talking to each other.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Manager pattern (agents as tools)
One agent stays in charge and calls specialist agents the way it would call any other tool.
Sources
- Anthropic, Building effective agents (2024-12-19)
- Anthropic, Multi-agent research system engineering post (2025-06-13)
- Google Research, Towards a science of scaling agent systems (2026-01-28)
- Anthropic, When to fold subagents into the main agent loop (2026-10-22)
- OpenAI, A practical guide to building agents (PDF)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.