dispatching-parallel-agents

name	dispatching-parallel-agents
description	Use when a task has multiple independent subtasks that can be executed concurrently by separate agents. Triggers when decomposed work has 2+ subtasks with no data dependencies, when subtasks operate on different files or codebase sections, when serial execution time would be significantly longer than parallel, or when independent analyses or deliverables need concurrent generation.

Dispatching Parallel Agents

Overview

This skill coordinates multiple agents working concurrently on independent subtasks to reduce total execution time while maintaining correctness. It provides strict rules for identifying safe parallelization opportunities, writing focused agent prompts, and integrating results without conflicts. The key constraint is that no two agents may modify the same file.

Announce at start: "I'm using the dispatching-parallel-agents skill to run [N] independent tasks concurrently."

Agent Tool Reference

All dispatch uses the Agent tool. Parameters:

• prompt (required) — task description with full context
• description (required) — short label (3-5 words)
• subagent_type — "Explore" (codebase search), "Plan" (architecture), "general-purpose" (default)
• run_in_background — true for async (you'll be notified on completion)
• model — optional override: "sonnet", "opus", "haiku"

Parallel: Multiple Agent calls in one message run concurrently. Background: run_in_background=true for non-blocking work. Named agents: Use subagent_type to reference installed agent templates (e.g., "superpowers:code-reviewer").

Trigger Conditions

• A task decomposes into 2+ subtasks with no data dependencies between them
• Each subtask operates on different files or different sections of the codebase
• The combined result can be assembled after all agents complete
• Total serial time would be significantly longer than parallel time
• /decompose output reveals independent task clusters

---

Phase 1: Independence Verification

Goal: Confirm subtasks are truly independent and safe to parallelize.

Every subtask must satisfy ALL four independence criteria:

Criterion	Question	If NO
No shared files	Do any two agents write to the same file?	Serialize those tasks
No shared mutable state	Does any agent depend on a side effect of another?	Serialize dependent tasks
Self-contained context	Can each agent work with only its own inputs?	Provide more context or serialize
Independent verification	Can each agent's output be validated alone?	Combine into single task

Parallelization Decision Table

Scenario	Parallelize?	Reason
Different files, different concerns	Yes	No conflict possible
Same module, different files	Yes (careful)	Verify no shared imports change
Same file, different sections	No	Merge conflicts inevitable
Task B uses Task A's output	No	Sequential dependency
Both read same files, write different	Yes	Reads are safe to parallelize
Both modify shared config file	No	Config conflicts
Independent test files	Yes	Tests are independent
One agent adds dep, another uses it	No	Package-level dependency

When NOT to Parallelize

• Subtasks share mutable state or modify the same files
• Task B depends on the output of Task A
• The overhead of coordination exceeds the time saved
• A single agent can complete the work in under 30 seconds
• The task requires iterative refinement where each step informs the next

STOP — Do NOT dispatch agents (via the Agent tool) until:

• All four independence criteria verified for every subtask pair
• No two agents write to the same file
• Each agent's context is self-contained

---

Phase 2: Prompt Construction

Goal: Write focused, unambiguous prompts that prevent scope creep and conflicts.

Each agent prompt MUST contain exactly four sections:

Section 1: Scope (What to Do)

Be specific about the exact task, files, and expected changes.

SCOPE: Add structured JSON logging to all API route handlers in src/api/.
Replace console.log calls with the logger from src/utils/logger.ts.
Files to modify: src/api/users.ts, src/api/orders.ts, src/api/products.ts.

Section 2: Context (Everything Needed)

Provide all information the agent needs without requiring it to explore.

CONTEXT:
- Logger API: logger.info(message, metadata), logger.error(message, error)
- Import: import { logger } from '../utils/logger'
- Current pattern in files: console.log('action', data)
- Target pattern: logger.info('action', { data, requestId: req.id })

Section 3: Output Format (What to Return)

Define exactly what the agent should produce.

OUTPUT: For each modified file, return:
1. The file path
2. A summary of changes made
3. Number of console.log calls replaced

Section 4: Constraints (What NOT to Do)

Prevent scope creep and conflicts explicitly.

CONSTRAINTS:
- Do NOT modify any files outside src/api/
- Do NOT change the logger utility itself
- Do NOT add new dependencies
- Do NOT refactor function signatures
- Do NOT modify test files
- If you encounter an issue outside your scope, report it but do not fix it

Agent Prompt Template

You are a focused agent with a single task.

## Scope
[Specific task description with exact files]

## Context
[All information needed to complete the task]
[Relevant code patterns, APIs, conventions]

## Output Format
[Exact structure of what to return]

## Constraints
- Do NOT modify files outside: [list]
- Do NOT change: [list things to leave alone]
- Do NOT add dependencies
- If you encounter an issue outside your scope, report it but do not fix it

Prompt Quality Checklist

Check	Question
Scope is specific	Can the agent complete the task without guessing?
Context is complete	Does the agent need to explore the codebase? (should be no)
Output is defined	Will the agent return what you need to integrate?
Constraints are explicit	Are file boundaries and "do NOT" items clear?

STOP — Do NOT dispatch (via the Agent tool) until:

• Every prompt has all 4 sections
• No prompt requires the agent to explore beyond provided context
• File boundaries are explicit in every constraint section

---

Phase 3: Dispatch and Monitor

Goal: Launch all agents (via the Agent tool) concurrently and track completion.

1. Launch all agents concurrently by invoking multiple Agent tool calls in a single message
2. Each agent works in isolation on its designated files
3. Monitor for completion — wait for ALL agents to finish
4. Collect outputs from every agent

Dispatch Tracking Table

| Agent | Task | Status | Files | Result |
|-------|------|--------|-------|--------|
| Agent 1 | Add logging to API | in_progress | src/api/*.ts | — |
| Agent 2 | Update unit tests | in_progress | tests/unit/*.ts | — |
| Agent 3 | Fix CSS layout | in_progress | src/styles/*.css | — |

Failure Handling During Dispatch

Scenario	Action
One agent fails, others succeed	Retry failed agent independently (via the Agent tool)
Multiple agents fail independently	Retry each independently (via the Agent tool)
Agent reports out-of-scope issue	Note for post-integration review
Agent exceeds scope (modifies wrong files)	Reject output, re-dispatch (via the Agent tool) with stricter constraints

---

Phase 4: Integration and Verification

Goal: Combine all agent outputs and verify the integrated result.

1. Collect outputs — Gather results from every agent
2. Check for conflicts — Verify no file was modified by multiple agents
3. Apply changes — Integrate all outputs into the codebase
4. Run integration checks — Execute the full test suite
5. Resolve issues — If integration fails, identify which agent's changes caused it
6. Commit atomically — All changes go in together or not at all

Integration Verification Checklist

Check	Command	Must Pass
No file conflicts	Diff outputs for shared files	Yes
Tests pass	Full test suite	Yes
Build passes	Build command	Yes
Lint passes	Lint command	Yes
No regressions	Compare test count before/after	Yes

Integration Failure Decision Table

Failure Type	Diagnosis	Action
Test failure in Agent 1's files	Agent 1's changes have a bug	Re-dispatch Agent 1 (via the Agent tool) with test failure context
Test failure in unrelated files	Cross-cutting regression	Identify root cause, fix manually or re-dispatch (via the Agent tool)
Build failure	Import/type issue	Check which agent's changes caused it, fix
Merge conflict	Agents touched same file (should not happen)	Rollback, serialize those tasks

STOP — Do NOT commit until:

• All agent outputs collected
• No file conflicts detected
• Full test suite passes
• Build and lint pass

---

Common Parallel Patterns

Pattern Decision Table

Pattern	When to Use	Example
By Module	Independent modules or packages	One Agent call per microservice
By Layer	Layers touch different files	API agent, service agent, data agent
By Feature Area	Independent vertical slices	Auth agent, profile agent, billing agent
By Task Type	Code, tests, docs touch different files	Code agent, test agent, docs agent

Example: Full Dispatch

TASK: "Update the API to v2, add tests, and update OpenAPI spec"

AGENT 1 - API Routes:
  Scope: Update route handlers in src/routes/v2/
  Context: [v2 API spec, breaking changes list]
  Output: Modified files list, breaking changes implemented
  Constraints: Do NOT touch tests or docs

AGENT 2 - Tests:
  Scope: Write tests in tests/v2/
  Context: [v2 API spec, test conventions, existing v1 tests as reference]
  Output: New test files, coverage summary
  Constraints: Do NOT modify source code

AGENT 3 - OpenAPI Spec:
  Scope: Update openapi/v2.yaml
  Context: [v2 API spec, OpenAPI 3.1 format]
  Output: Updated spec file
  Constraints: Do NOT modify code or tests

---

Anti-Patterns / Common Mistakes

Anti-Pattern	Why It Fails	Correct Approach
Two agents modifying the same file	Merge conflicts, data loss	One file owner per Agent dispatch
Shared mutable state between agents	Race conditions, inconsistency	Eliminate shared state
Incomplete context in prompts	Agents explore and step on each other	Provide ALL needed context
Vague file boundaries	Agents guess scope, modify wrong files	Explicit file lists in constraints
No integration check after completion	Cross-cutting bugs go undetected	Full test suite after integration
Parallelizing sequential tasks	Agent B needs Agent A's output	Verify independence first
Not tracking which agent touched which file	Cannot diagnose integration failures	Maintain dispatch tracking table
Dispatching too many agents (10+)	Coordination overhead exceeds savings	2-5 Agent calls per dispatch round
Skipping rollback preparation	Cannot recover from integration failure	Keep pre-dispatch state recoverable

---

Anti-Rationalization Guards

Do NOT dispatch agents (via the `Agent` tool) that modify the same file. Do NOT parallelize tasks with data dependencies. Do NOT skip integration verification. If independence criteria are not met, serialize the tasks.

If you catch yourself thinking:

• "These agents probably won't conflict..." — Verify. Do not assume.
• "The integration will be fine..." — Run the full test suite. Always.
• "I can merge their changes to the same file manually..." — No. One file, one owner.

---

Integration Points

Skill	Relationship	When
task-decomposition	Upstream — identifies independent task clusters	Before dispatching
subagent-driven-development	Complementary — provides review gates	Quality gates for agent output
executing-plans	Upstream — may delegate independent tasks	During plan execution
verification-before-completion	Downstream — verifies integrated result	After integration
code-review	Downstream — reviews integrated changes	After all agents complete
resilient-execution	On failure — retries failed agents	When individual agents fail

---

Parallelism Safety Rules Summary

Rule	Rationale
No two agents modify the same file	Prevents merge conflicts and race conditions
No shared mutable state	Eliminates data races
Each agent gets complete context	Prevents agents from exploring and stepping on each other
Define file boundaries explicitly	Makes ownership unambiguous
Review integration after completion	Catches cross-cutting issues
Atomic commit for all changes	All in or all out
Always have a rollback path	Keep pre-dispatch state recoverable

---

Skill Type

RIGID — Follow this process exactly. Independence verification is mandatory. All four prompt sections are mandatory. Integration verification is mandatory. No shortcuts on parallelism safety.