A subagent handles a bounded task within a larger agent-driven process. It keeps its own context window, instructions, tool calls, observations, and reasoning path; a coordinator provides limited input and receives a compact result, not the full trajectory. section 7.5.2
Design the boundary around one concern:
- Give the subagent a narrow task, minimum necessary input, limited tools, and an explicit output contract. section 7.5.2
- Order dependent subtasks explicitly to avoid inconsistent intermediate states; account for the added coordination cost. section 7.5.2
- Use subagents to isolate context; state partitioning selects fields for prompts, while environment sandboxing keeps runtime state outside prompt memory. section 7.5.2






