parallel-dispatch

name	parallel-dispatch
model	opus
description	Produces a verified multi-track execution by dispatching parallel agents with file manifests, independent verification, and status tracking. Use when: "dispatch agents", "parallel tracks", "multi-track work", "fan out", "divide and conquer".
category	agent-orchestration
triggers	["dispatch agents in parallel","parallel tracks","multi-track work","fan out","divide and conquer"]
tier	1
agents	["primary"]
tool_dependencies	["file_system"]
inputs	[{"name":"tracks","type":"array","description":"List of track descriptions, each with objective, scope, and expected output","required":true}]
outputs	[{"name":"dispatch_report","type":"text","format":"markdown-table","description":"Verified status table with per-track agent status, verification result, and files changed"}]

Parallel Dispatch Skill

Philosophy

Parallel dispatch is not "fire and forget" — it is "fire, verify, and integrate." The value of parallelism comes from independent execution, but the risk comes from unverified output. Every agent is treated as an untrusted contractor: given a clear scope, expected to deliver, but verified independently before their work is accepted. The dispatch orchestrator's primary job is not delegation — it is verification.

When to Use

• When 2+ independent tracks can execute concurrently without file conflicts
• For multi-module fixes, refactors, or feature additions across a workspace
• When time pressure demands parallel execution over sequential safety
• For commission-style work where each track has clear boundaries

Workflow

Phase 1: Plan

1. Parse tracks from user input
2. For each track, define:
   • Clear objective and scope
   • Specific file manifest (files the agent may touch)
   • Expected outputs and success criteria
   • Build/test commands to verify
3. Check for file conflicts — if two tracks touch the same file, use isolation: "worktree" or serialize them
4. Create a TodoWrite checklist with one item per track

Phase 2: Dispatch

1. Dispatch agents in parallel — one per track, max 5 concurrent
2. Each agent prompt must include:
   • The full file manifest and success criteria
   • Instruction: "Run all tests before reporting completion"
   • The build/test commands that will be used to verify
3. Model routing: Sonnet for straightforward implementation, Opus for architectural/complex tracks

File Manifest Protocol

Every agent prompt must include an explicit file manifest stating exactly which files the agent may create or modify. This is the zero-overlap guarantee that makes parallel execution safe.

Agent Dispatch Template:

"You own ONLY these files: [explicit list]. DO NOT read, modify, or create any files outside this list. If you need changes to files outside your manifest, report the need in your output — do not make the change."

Ownership rules:

• Agents that CREATE new files never conflict — new-file creation is inherently exclusive
• Agents that MODIFY existing files need exclusive ownership — one agent per file, no exceptions
• One agent per concern, not per file. Assign a complete feature slice (component + test + wiring), not a single file
• Main thread integrates: after all agents land, run , , wire cross-component imports, then one final build check