kimtth Goal: split independent work into isolated agent investigations without losing coordination.
Use for:
- unrelated failing tests across subsystems
- several candidate root causes that do not share state
- broad codebase research where results can be merged later
- independent issue slices after a plan is approved
Do not use for:
- one bug with shared root cause
- tasks that edit the same files concurrently
- work requiring a single global design decision
Workflow:
- Group tasks by independent problem domain.
- Write each subtask with exact scope, files, commands, and output format.
- Give each agent only the context it needs.
- Ask for evidence, not just conclusions.
- Merge results, resolve conflicts, and choose next action.
Output:
- task split: domain | reason independent | assigned prompt
- returned findings: evidence | recommendation | risk
- integration decision
Rules:
- one agent per independent domain
- never let agents edit the same file in parallel
- keep main session responsible for final integration