subagent-driven-development

Load the plan into a TodoWrite list, then dispatch a fresh general-purpose subagent for each task with a focused prompt. After each task, dispatch a code-reviewer subagent with the base and head commit SHAs, fix any Critical or Important issues, then proceed to the next task.

Use sequential execution when tasks are tightly coupled or must run in order. Use parallel execution when tasks are independent—different files, subsystems, or bugs—so agents can work concurrently without conflicts, with review after each batch.

Group failures by independent problem domain, such as one test file or subsystem per agent. Give each agent a focused scope, the exact error messages, constraints on what not to change, and a required summary output, then run the full test suite after integration.

Avoid it when failures are related and fixing one may fix others, when understanding requires full system context, during exploratory debugging, or when agents would share state such as editing the same files or resources.

A good prompt is focused on one problem domain, self-contained with all needed context including error messages, explicit about constraints like not changing production code, and specific about the expected output such as a summary of root cause and changes.