Dispatching parallel agents

Gate (required)

Use this skill only if:

• User or parent explicitly allows parallel subagents, and
• Tasks are independent (no shared files, no ordering dependency on partial output)

If either condition fails, use sequential incremental-implementation instead.

When to use

• Independent reviews (security + tests + API contract) on the same commit
• Parallel research spikes with separate output files
• Large plan with disjoint file ownership per task

When not to use

• Single shared module being edited by multiple agents
• Database migrations or sequential deploy steps
• User did not approve multi-agent execution

Workflow

1. Partition — list tasks with disjoint Files: sets
2. Brief — each agent gets scope, constraints, verify command, output format
3. Dispatch — run parallel agents; cap concurrency to what the harness supports
4. Collect — gather outputs without assuming correctness
5. Integrate — parent reconciles conflicts, runs full verification, single commit strategy per user rules

Integration checklist

• No conflicting edits to the same paths
• Combined diff reviewed by parent (or requesting-code-review)
• Full test/lint suite on integrated result
• User informed of what each agent did

Anti-patterns

• Parallel implementers on the same feature branch without file boundaries
• Merging subagent output without reading diffs
• Spawning agents to avoid reading the codebase

Related

subagent-driven-development, incremental-implementation, requesting-code-review, verify-before-done

Orchestration patterns inspired by obra/superpowers (MIT).