name: dispatch-parallel-agents
description: Dispatch independent tasks to parallel subagents when 2+ problems can be solved without shared state or sequential dependencies.
license: MIT
compatibility:
opencode: ">=0.1"
metadata:
category: execution
phase: implementation
Skill: Dispatch Parallel Agents
This skill parallelizes independent work across multiple subagents. Each agent gets a focused scope, isolated context, and a clear deliverable.
When to Use
Use when:
- 2+ independent tasks exist that do not share state
- Each task can be understood and completed without context from the others
- Tasks do not edit the same files
Do not use when:
- Failures are related (fixing one might fix others -- investigate together first)
- Tasks require understanding the full system state
- Agents would interfere with each other (editing the same files, using the same resources)
- You are in exploratory mode and do not yet know what is broken
Execution Model
Roles
- Primary (coordinator)
- Identifies independent problem domains
- Crafts focused prompts for each agent
- Reviews results and integrates changes
- Resolves conflicts if agents touched overlapping code
- Subagents (workers)
- Each receives a scoped task with all necessary context
- Works independently without knowledge of other agents
- Returns a summary of findings and changes
Workflow
Step 1: Identify Independent Domains
Group tasks by what they affect. Each domain must be independent -- fixing one must not affect the others.
Example:
- Domain A: Fix stats page header styling
- Domain B: Update billing page layout
- Domain C: Add admin role detection
These are independent because they touch different files and different logic.
Step 2: Craft Agent Prompts
Each agent prompt must be:
- Focused -- one clear problem domain, not "fix everything"
- Self-contained -- all context needed to understand the problem (file paths, error messages, expected behavior)
- Constrained -- explicit boundaries on what the agent should and should not change
- Output-specific -- what the agent should return (summary, file list, verification result)
Step 3: Dispatch
Launch all agents in a single message with multiple Agent tool calls. This ensures true parallel execution.
Step 4: Review and Integrate
When agents return:
- Read each agent's summary
- Check for conflicts (did agents edit the same code?)
- Verify changes work together (run build, run tests)
- Integrate all changes
Rules
- Independence is mandatory: If tasks share state or files, do not parallelize. Process sequentially instead.
- Focused prompts: Each agent gets exactly what it needs. Do not dump full session context into agent prompts.
- Verify after integration: Always run the build and relevant tests after merging all agent outputs. Agents cannot verify cross-agent interactions.
- Do not over-parallelize: 2-3 agents is typical. More than 5 agents indicates the tasks should be structured differently (e.g., as a plan with phases).
- Conflicts require manual resolution: If two agents edited the same file, the primary resolves the conflict -- do not blindly apply both changes.