One agent per independent problem domain, all dispatched in a single response so they run concurrently. Without subagent dispatch, run the tasks one at a time in this session; independence is the guarantee, parallelism is only the cheap way to get it.

Rules:

Each agent's prompt carries scope, context, constraints, and expected return. Scope: one file, one subsystem, one problem. "Fix all the tests" loses the agent. Context, pasted in: error messages, test names, relevant paths. Agents inherit nothing from the session. Constraints: what must not change ("tests only", "do not touch production code"). Expected return: a summary of root cause and changes, so integration is reviewable.
Integrate when they return: read each summary, check the diffs for conflicts, then run the verify skill, whose full gate covers the suite, before claiming the batch done.
Do not use when failures are related (fixing one may fix the rest: investigate together first), when the problem is still exploratory, or when agents would touch the same files.
A batch is the biggest slice you did not write: offer the user /recall on the integrated batch.