
Orchestrator-worker pattern: one lead model delegating
In short: In the orchestrator-worker pattern one LLM splits a task into subtasks at run time, delegates each to a worker call and merges the results into one answer.
The orchestrator-worker pattern is a design where one model, the orchestrator, reads the task, decides which subtasks are needed, sends each to a worker model call or subagent, and combines the results. Anthropic's December 2024 guide describes a central LLM that dynamically breaks down tasks, delegates them to worker LLMs and synthesizes their results. It suits jobs where the shape of the work depends on the input.
Orchestrator-worker pattern in plain words
Picture a project lead who reads the brief, decides "this needs legal, design and a data pull", hands those out, then stitches the answers into one reply.
How it works
- The orchestrator analyses the input and produces a list of subtasks with instructions.
- Workers run, often in parallel, each with a narrow prompt and only the context it needs.
- Workers return compact results.
- The orchestrator synthesises, checks for gaps and may issue a second round.
Anthropic's account of its multi-agent research system, published in June 2025, shows the pattern at work: a lead agent spins up three to five subagents in parallel, each returns condensed findings, and the lead decides whether more research is needed.
Example
As an illustration, take a pull request touching five services. The orchestrator decides which files changed meaningfully, assigns one worker per service to review the change against that service's conventions, then writes a single review comment ranked by severity.
Where you'll meet it
- Coding agents planning multi-file changes
- Deep-research tools that fan out searches
- Document review across many sections or contracts
- Agent frameworks such as LangGraph, whose documentation lists orchestrator-worker among its patterns
What it changes for you
It lets a system take on bigger jobs without one giant context. For developers, the orchestrator's planning prompt and the worker result format are the two things to design carefully. For finance, cost scales with worker count: the same Anthropic write-up reports that its multi-agent systems used about 15 times more tokens than chat interactions. Cap the number of workers.
Orchestrator-worker vs parallelisation workflow
A parallelisation workflow runs subtasks you fixed in code. An orchestrator decides the subtasks itself, per input.
Common mistake
My view: letting workers return long essays. The orchestrator then drowns in its own inputs. Make workers return a fixed, short format, such as findings with file and line, and the merge step becomes reliable.
Related terms
Multi-agent system · Subagent · Agentic workflow · Context window · Cost per task
From the archive: Agent swarms need aggregation evals
Make every worker return a short, fixed format so the orchestrator can merge the results reliably.


