Orchestrator-Worker
also called Lead agent, Supervisor pattern
One agent plans and delegates; others do the work.
A pattern where a lead agent decomposes a task, delegates independent parts to worker agents or subagents, then synthesises their results. The orchestrator holds the plan and the workers hold narrow, disposable context.
It earns its cost when subtasks are genuinely independent and each returns a small summary. When they are not, the coordination overhead and the reconciliation of conflicting work usually cost more than doing it in one well-managed context.
The published production accounts are consistent about where the difficulty lives: not in spawning subagents, but in describing subtasks precisely, deciding how much each should return, and evaluating a result assembled from several sources. That last part is where the quality control has to live.
Parallel research with a token budget per worker
Investigating a question across many independent sources.
Give each worker a narrow sub-question and a cap on what it may return. The orchestrator synthesises brief findings rather than raw material, which is both cheaper and more reliable than letting each worker return everything it found.
When one agent is the better answer
A task where the subtasks all touch the same three files.
Workers will conflict, duplicate edits, and require reconciliation. When the work is not genuinely independent, one agent with a well-managed context is simpler, cheaper and less likely to produce something inconsistent.
Adopting it because it sounds scalable. Orchestration adds latency, cost and a new class of coordination failure; use it when the parallelism is real.
- blogBuilding effective agents ↗
Draws the line between fixed workflows and agents that choose their own process, and names the core patterns: chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer.
- blogHow we built our multi-agent research system ↗
Production account of a lead-agent plus subagents architecture, including when delegation helps and what the coordination costs are.
Subagent
A helper agent with its own clean context window.
Harness · Scaffold
Everything around the model that makes it an agent.
Agent
A model given tools, a goal and permission to keep going on its own.
Handoff
Passing work to a fresh context, deliberately and with a written brief.
Spec · Specification
A written description of the intended outcome, before the work starts.