Multi-agent work needs a preflight check
Several agents are justified only when parallel benefit exceeds coordination cost. Default to one agent, especially for a small or sequential task.
Independence comes from boundaries and contracts, not from the number of role names. Two agents changing one module are not truly independent.
| Good fit | Counter-criterion |
|---|---|
| Backend, frontend, and tests have separate paths | Small local change in one file |
| Read-only agents test different hypotheses | Reproduce → fix → test must stay sequential |
| Reviewers have distinct focuses | Several agents edit one module |
| Each role has its own success check | Public contract is not defined |
| There is time for a human checkpoint | No tests or safety net |
Safer escalation path
- Try one session or plan mode.
- Add one read-only investigator for broad search.
- Add an independent reviewer when a second opinion is useful.
- Define contracts and isolated paths before parallel edits.
- Use several executors only after integration and ownership are explicit.
Decision note
task: <id>
size: large
workstreams_independent: yes
ownership_split_by: modules
tests_available: yes
shared_state_risk: low
verdict: multi-agent
human_checkpoint: before mergeCanonical lesson source · JavaRush
Open JavaRush material