Aller au contenu principal

Skill: parallel-agents

Fork

Orchestration of parallel agents to maximize efficiency. Trigger when a task can be decomposed into independent sub-tasks that can run in parallel.

Configuration

PropertyValue
Contextfork
Allowed toolsRead, Glob, Grep, Bash
Keywordsparallel, agents

Detailed description

Parallel Agents Orchestration

Goal

Decompose complex tasks into independent sub-tasks and run them in parallel via specialized sub-agents to maximize efficiency.

When to use parallelism

┌──────────────────────────────────────────────────────────────────┐
│ DECISION: PARALLEL OR SEQUENTIAL ? │
├──────────────────────────────────────────────────────────────────┤
│ │
│ PARALLEL if: │
│ - INDEPENDENT sub-tasks (no data dependency) │
│ - MERGEABLE results (combinable without conflict) │
│ - Task DECOMPOSABLE into distinct parts │
│ │
│ SEQUENTIAL if: │
│ - Result A required to start B │
│ - Modifications on the SAME files │
│ - Execution order MATTERS │
│ │
└──────────────────────────────────────────────────────────────────┘

Parallelization patterns

1. Fan-Out / Fan-In

┌─→ [Agent A: security audit] ─→┐
│ │
[Task] ─┼─→ [Agent B: perf audit] ─→┼─→ [Combined report]
│ │
└─→ [Agent C: a11y audit] ─→┘

Usage: Audits, multi-criteria analyses, parallel reviews

2. Map-Reduce

[Files] ─→ [Agent 1: file A] ─→┐
→ [Agent 2: file B] ─→┼─→ [Synthesis]
→ [Agent 3: file C] ─→┘

Usage: Code analysis by module, tests by domain

3. Pipeline with parallel steps

[Step 1] ─→ [Step 2a] ─→┐
[Step 2b] ─→┼─→ [Step 3]
[Step 2c] ─→┘

Usage: Build pipeline, workflow with independent steps

Common parallelizable tasks

Audits and analyses

TaskParallel agentsResult
Full auditqa-security + qa-perf + wcag-auditCombined report
Code reviewqa-review per module/fileIssue list
Explorationwork-explore per functional domainCode map

Development

TaskParallel agentsResult
Tests per moduledev-test per serviceTest suite
Documentationdoc-generate per componentComplete docs
Migrationops-migrate per dependencyComplete migration

Business

TaskParallel agentsResult
Market researchbiz-competitor + biz-personasComplete analysis
Launchgrowth-landing + growth-seo + growth-analyticsLaunch kit

How to dispatch

Step 1: Decompose the task

## Main task: [Description]

### Identified sub-tasks:
1. [ ] [Sub-task A] - Agent: [type] - Independent: Yes/No
2. [ ] [Sub-task B] - Agent: [type] - Independent: Yes/No
3. [ ] [Sub-task C] - Agent: [type] - Independent: Yes/No

### Dependencies:
- A → independent
- B → independent
- C → depends on A and B

### Plan:
- Phase 1 (parallel): A + B
- Phase 2 (sequential): C (after A and B)

Step 2: Launch in parallel

Use the Task tool with multiple calls in a single message:

[Call 1] Task(subagent_type="qa-security", prompt="Audit...")
[Call 2] Task(subagent_type="qa-perf", prompt="Analyze...")
[Call 3] Task(subagent_type="wcag-audit", prompt="Verify...")

Step 3: Combine results

## Combined report

### Agent A: [Summarized results]
### Agent B: [Summarized results]
### Agent C: [Summarized results]

### Synthesis
[Overview and priorities]

File conflict prevention

IMPORTANT: Parallel agents editing the same files cause race conditions and broken builds.

Before parallelizing, establish a file map

### Files per agent:
- Agent A: src/auth/ (exclusive)
- Agent B: src/api/ (exclusive)
- Agent C: src/utils/helpers.ts (CONFLICT with A and B!)

→ Solution: Agent C sequential after A and B

File-locking rules

SituationAction
2 agents modify the same fileSEQUENTIAL mandatory
2 agents modify the same folderCheck the specific files
Read-only agents (audit)PARALLEL always OK
Shared config (package.json, tsconfig)SEQUENTIAL for edits

Typically shared files (caution)

  • package.json — deps added by multiple agents
  • tsconfig.json — paths modified
  • src/index.ts — exports added
  • .env.example — variables added
  • Routing/navigation files

Best practices

  • Verify the independence of sub-tasks BEFORE parallelizing
  • Establish the map of files modified per agent BEFORE launching
  • Give each agent a clear and bounded scope
  • Use run_in_background: true for long tasks
  • Combine results with a high-level synthesis
  • Limit to 3-5 parallel agents for readability
  • Prefer isolation: "worktree" for agents that edit many files

For complex orchestrations requiring inter-agent communication, prefer native Agent Teams:

Sub-Agents (Task)Agent Teams
CommunicationReturn to parent onlyDirect messaging between agents
CoordinationMain agent handles everythingShared task list
Token costLowHigh (1 context per agent)
Ideal forFocused tasks, combined resultsComplex collaboration, debate, consensus

Recommendation: Use Task sub-agents (this skill) for focused and independent tasks. Use Agent Teams (/work:work-team) for teams of 3+ agents requiring discussion and coordination.

See the agent-teams skill for full documentation.

Rules

  • ALWAYS verify dependencies between sub-tasks
  • NEVER parallelize modifications on the same files
  • ALWAYS establish the map of modified files BEFORE launching agents
  • ALWAYS provide full context to each agent
  • COMBINE results into a coherent report
  • PREFER isolation: "worktree" when agents modify many files

Automatic triggering

This skill is automatically activated when:

  • The matching keywords are detected in the conversation
  • The task context matches the skill's domain

Triggering examples

  • "I want to parallel..."
  • "I want to agents..."

Context fork

Fork means the skill runs in an isolated context:

  • Does not pollute the main conversation
  • Results are returned cleanly
  • Ideal for autonomous tasks

See also