Dispatching parallel agents

Dispatching Parallel Agents

Quick Reference

When you have 2+ truly independent tasks, launch them as parallel agents in a single message. You orchestrate the fan-out and fan-in. Keep the main context clean.

When to Use

• 2+ research questions that don't depend on each other
• Independent code investigations across different parts of the codebase
• Multiple review tasks (security, performance, patterns) on the same diff
• During plan execution when independent steps are identified

When NOT to Use

• Tasks have dependencies (agent B needs agent A's result)
• Single task that's just large (use one focused agent instead)
• Tasks that need conversational context (agents start fresh)
• The overhead of dispatching exceeds just doing it sequentially

Core Process

Step 1: Identify Independent Work

Tasks are independent when:

• Agent A's result doesn't change agent B's task
• They don't modify the same files
• They don't need to coordinate
• Order doesn't matter

Step 2: Write Agent Briefs

Each agent gets a self-contained prompt:

You are tasked with [specific goal].

Context:
- [Relevant background]
- [Files to look at]

Task:
- [Concrete steps]

Output:
- [What to return — findings, code, recommendations]

Constraints:
- [What NOT to do]
- [Read-only? Implementation? Research only?]

Be explicit about whether the agent should:

• Research only: Read, search, analyze — return findings
• Implement: Write code, create files — return what was changed
• Review: Analyze code quality — return findings

Step 3: Dispatch in Parallel

Launch ALL independent agents in a single message using multiple Agent tool calls:

[Agent 1: Research authentication patterns]
[Agent 2: Research database migration options]
[Agent 3: Review existing API structure]

Use appropriate agent types:

• Explore — for codebase exploration and research
• general-purpose — for implementation or complex multi-step tasks
• Specialized types — for domain-specific reviews

Step 4: Fan-In

After all agents return:

1. Summarize each agent's findings
2. Look for conflicts or contradictions between agents
3. Synthesize a unified recommendation or proceed with implementation

Step 5: Decide Next Action

Based on agent results:

• All research complete → proceed with plan/implementation
• Implementation complete → chain to verification
• Contradictions found → resolve, potentially re-dispatch

Dispatch Patterns

Research Fan-Out

Agent 1: "How does auth work in this codebase?"
Agent 2: "What database patterns are used?"
Agent 3: "What test patterns exist?"
→ Synthesize into a plan

Review Fan-Out

Agent 1: "Security review of this diff"
Agent 2: "Performance review of this diff"
Agent 3: "Pattern compliance review of this diff"
→ Merge findings, prioritize by severity

Implementation Fan-Out

Agent 1: "Implement component A in worktree"
Agent 2: "Implement component B in worktree"
→ Merge worktrees, integration test

Anti-Patterns

Anti-Pattern	Why It's Wrong
Dispatching dependent tasks in parallel	Agent B will miss Agent A's output
Launching agents sequentially when they could be parallel	Wastes time. Use one message, multiple tool calls.
Duplicating work between main context and agents	If you delegate, don't also do it yourself
Too many agents at once (>5)	Diminishing returns, harder to synthesize
Vague prompts	Agents need concrete, self-contained tasks
Not synthesizing results	Raw agent output isn't useful. Summarize and decide.

Chaining

OPTIONAL: syntaxninja-dojo:verification-before-completion (after implementation agents complete) OPTIONAL: syntaxninja-dojo:subagent-driven-development (for implementation-focused parallel work)