Learn · Agentic architecture & orchestration · CCAR-F task 1.2

Coordinators and subagents

A coordinator splits the work, subagents dig in their own context windows, and only summaries come back. Hub-and-spoke, by design.

Key ideas

• A coordinator breaks the task down, delegates to subagents and synthesises the results.
• Context isolation: each subagent has its own context window; only its summary returns.
• Specialisation: each subagent can have its own system prompt, tools and model.
• Hub-and-spoke: all communication, routing and error handling go through the coordinator.
• Select subagents by query complexity, and partition scope so they don't duplicate work.
• Refine iteratively: check the draft for gaps, re-delegate targeted searches, re-synthesise.
• Costs are more tokens, coordination overhead and information loss in summaries.

Worked example: A due-diligence coordinator

Asked to assess a supplier, the coordinator sends one subagent to financial filings and another to news and litigation, so their sources don't overlap. The draft shows nothing on data security, so it sends a third, targeted brief before writing the final report.

Exam traps

• Letting subagents message each other directly instead of routing through the coordinator.
• Running the full multi-agent pipeline for a simple question.
• Assuming subagents are free: they use more total tokens and lose detail in summaries.
• Overlapping subagent scopes, so two of them search the same sources.

Official docs

• Subagents in Claude Code
• Subagents in the Agent SDK
• Building effective agents

Practise CCAR-F questions · Next lesson: Configuring subagents