dispatching-parallel-agents

<!-- compact: SKILL-compact.md — if it exists in this directory, load it instead of following the full instructions below. The compact version contains the same parallel dispatch pattern and orchestration patterns. -->

Dispatching Parallel Agents

Core principle: One agent per independent problem domain. Let them work concurrently.

When to Use / Not Use

Use: 3+ failures/tasks with different root causes, no shared state, each problem self-contained. Skip: Failures are related (fix one might fix others), agents would edit same files, you don't know what's broken yet.

The Pattern

1. Identify Independent Domains

Group by what's broken. E.g.: File A = tool approval flow, File B = batch completion, File C = abort. Each is independent.

2. Focused Agent Tasks

Each agent gets: specific scope, clear goal, constraints (don't touch other code), expected output format.

3. Dispatch in Parallel

Every spawn names a model — execution model: sonnet, validation model: opus, frontier-in-the-loop model: inherit (routing doctrine, commands/model-routing.md). Unpinned spawns inherit the session model.

Task("Fix agent-tool-abort.test.ts failures", model="sonnet")
Task("Fix batch-completion-behavior.test.ts failures", model="sonnet")
Task("Fix tool-approval-race-conditions.test.ts failures", model="sonnet")
// All three run concurrently

4. Review and Integrate

Read each summary → verify fixes don't conflict → run full test suite.

Agent Prompt Structure

Fix the 3 failing tests in src/agents/agent-tool-abort.test.ts:

1. "should abort tool with partial output capture" - expects 'interrupted at' in message
2. "should handle mixed completed and aborted tools" - fast tool aborted instead of completed
3. "should properly track pendingToolCount" - expects 3 results but gets 0

These are timing/race condition issues. Your task:
1. Read the test file and understand what each test verifies
2. Identify root cause - timing issues or actual bugs?
3. Fix (replace arbitrary timeouts with event-based waiting, fix bugs, adjust expectations)

Do NOT just increase timeouts.
Return: Summary of what you found and what you fixed.

Prompt Mistakes

• Too broad: "Fix all the tests" → agent gets lost. Use: "Fix agent-tool-abort.test.ts"
• No context: "Fix the race condition" → agent doesn't know where. Paste error messages.
• No constraints: Agent might refactor everything. Add: "Do NOT change production code"
• Vague output: "Fix it" → you don't know what changed. Use: "Return summary of root cause and changes"

Verification

After agents return: review each summary → check for same-file conflicts → run full suite → spot check for systematic errors.

Cross-AI Variant: Codex Agents

Use Codex (via clavain:interserve): Well-scoped tasks with clear file lists, true parallel sandboxes, cost/context optimization. Keep Claude: Deep cross-file understanding, exploratory investigation, architectural decisions.

Orchestration Patterns

Parallel Specialists

Independent tasks, distinct domains, no shared files. All dispatched in one message.

Task("Review authentication module", model="sonnet")   ─┐
Task("Review database migrations", model="sonnet")     ─┤── all in one message
Task("Review API error handling", model="sonnet")      ─┘

Use when plan has 3+ independent modules.

Pipeline

Sequential handoff — agent N's output feeds agent N+1.

Agent 1: Research & design    → design.md
Agent 2: Implement from design.md → code changes
Agent 3: Write tests for code changes → test files

Use when tasks have strict data dependencies.

Fan-out / Fan-in

N agents, same question, different perspectives → synthesize.

Task("Review from security perspective", model="sonnet")     ─┐
Task("Review from performance perspective", model="sonnet")  ─┤── fan-out
Task("Review from UX perspective", model="sonnet")           ─┘
                     │
              Synthesize findings              ── fan-in

Used by /flux-drive and /quality-gates.