skill-dispatching-parallel-agents

Group failures by independent problem domain, then dispatch one focused subagent per test file or subsystem in the same round. Each agent gets specific scope, constraints, and a required output summary, and you review and integrate each lane's results afterward.

Use parallel dispatch when failures are independent, such as three or more test files failing with different root causes, and no shared state exists between investigations. Avoid it when failures are related, when full system context is needed, or when agents would edit the same files.

No, same-workspace concurrent writers remain prohibited. Parallel implementation lanes require an approved plan, isolated Git worktrees, disjoint write ownership, and declared lane proof; a read-only task may overlap one writer only with immutable inputs.

Good prompts are focused on one problem domain, self-contained with all needed context such as error messages and test names, and specific about constraints and expected output. Avoid broad instructions like "fix all the tests" or vague outputs like "fix it".

It does not fit related failures where fixing one might fix others, exploratory debugging, or work requiring full system state. If the platform lacks concurrent agent support, lanes must run sequentially rather than simulating parallelism with overlapping edits.