dispatching-parallel-agents

Group failures by independent problem domain, then dispatch one agent per domain with a focused prompt containing the failing test names, error context, constraints, and required summary output. After agents return, review each summary, check for conflicting edits, and run the full test suite.

Use parallel agents when failures are independent, each problem is understandable without other problems' context, and no shared state exists between investigations. If fixing one failure might fix others, or agents would edit the same files, debug sequentially instead.

A good agent prompt is focused on one problem domain, self-contained with all needed context such as error messages and test names, explicit about constraints like not modifying unrelated code, and clear about the expected output such as a summary of root cause and fixes.

Common mistakes include overly broad scopes like fixing all tests at once, missing context so agents cannot locate the problem, absent constraints letting agents refactor unrelated code, and vague output requirements that hide what was actually changed.

Avoid it for correlated failures where one fix may resolve others, problems requiring full system context, exploratory debugging where the cause is unknown, and tasks with shared state where agents would interfere by editing the same files or resources.