📝Orchestrator-Worker is a multi-agent design pattern in which one lead agent plans a task, delegates subtasks to worker agents, and combines their results into a final output.
The orchestrator holds the overall goal. It decomposes the work, writes a brief for each subtask, dispatches workers, and synthesizes what they return. Workers receive only their piece and a fresh context, so they stay focused and the orchestrator's own context stays free of their intermediate steps.
The pattern suits tasks that cannot be fully planned in advance, such as research across many sources or changes spanning many files, because the orchestrator can adjust the plan as results arrive. That distinguishes it from a fixed pipeline, whose steps are set beforehand. 📝Anthropic has described using this design for its multi-agent research system.
In 📝Claude Code the pattern appears as 📝Subagents dispatched from a main session and as 📝Claude Code Dynamic Workflows that orchestrate many subagents at once. Its main risks are vague briefs, which leave workers guessing, and token cost, which grows with each worker launched.