Vol. IV · Multi-Agent Architectures · IV.1
Orchestrator–Workers
A central model decomposes work it cannot predict in advance, delegates to workers, and synthesizes. The multi-agent pattern Anthropic actually endorses.
7 min read
Doctrine
- Workers get a fresh context plus a written brief, never the orchestrator's entire brain.
- The orchestrator should have tools like spawn_worker and read_file, not the workers' tools.
- Synthesis is a job. Budget time for it. Do not concatenate dumps and call it a report.
The shape
Anthropic: the orchestrator breaks down tasks dynamically — coding across files, research across sources. Deep Agents and DeerFlow implement it with subagents and a filesystem. LangGraph supervisor patterns encode it as a graph. CrewAI hierarchical process is the same idea with roles.
The hard part is the contract: what the worker may write (a file, a structured result), how failure is reported, and how the orchestrator decides it is done.
Anti-patterns
- Workers that can spawn more workers without a depth cap (fork bombs).
- A shared blackboard that every worker pollutes.