The simplest parallel pattern. Spawn N agents, each handling an independent slice, then synthesize results.
How it works:
- Break work into independent units (by file, module, or concern)
- Spawn one agent per unit with
run_in_background: true
- Each agent works autonomously and returns results
- The parent synthesizes when all complete
Best for: research across multiple areas, independent file analysis, running tests in parallel.
Key constraint: tasks must have no shared state and clear file boundaries. If agents edit overlapping files without worktree isolation, they’ll overwrite each other.
Context budget: Keep the orchestrator at ~10-15% of its context window. Have subagents write outputs directly to disk and return only a brief confirmation — this gives each subagent a fresh, full context window with no accumulated noise.
Examples in the wild
| Example | What it shows |
|---|
| planning-coordinator | 4 parallel researcher agents (code-explorer, arch-researcher, database-analyst, security-analyst) fan into a single plan. Explicit file ownership per agent. |
| deep-research | Breaks a topic into sub-questions, one agent per sub-question, synthesizes into a report. Write-to-disk / return-confirmation-only in practice. |
