dispatching-parallel-agents

name	dispatching-parallel-agents
description	Dispatch parallel subagents for independent tasks without shared state. Use when facing 2+ independent failures, debugging unrelated subsystems, or researching separate topics concurrently.

dispatching-parallel-agents

Dispatch parallel subagents when you have multiple independent tasks that don't share state. This skill provides the decision framework and prompt structure for effective parallelization.

When to Use

Parallel dispatch is appropriate when:

• Independent domains — Tasks touch different files, services, or subsystems with no overlap
• No shared state — No task depends on the output of another
• Clear boundaries — Each task has a well-defined scope and expected output
• Time savings — Running sequentially would be noticeably slower

When NOT to Use

Stay sequential when:

• Tasks depend on each other's output (chain, don't parallelize)
• Tasks modify the same files (merge conflicts are likely)
• The problem is not yet understood (investigate first, parallelize second)
• A single task is fast enough that parallelization adds overhead without benefit

The Pattern

Step 1: Identify Independent Tasks

Break the work into units that have no dependencies on each other. Each unit should be completable without knowledge of the others' results.

Questions to validate independence:

• Can task A complete without task B's output? (If no, they're dependent)
• Do tasks A and B modify the same files? (If yes, they'll conflict)
• Does the order matter? (If yes, they're sequential)

Step 2: Create Task Descriptions

Write a focused, self-contained prompt for each agent. Every prompt must include:

1. Context — What the agent needs to know about the codebase
2. Scope — Exactly what files/areas to work in
3. Objective — What to accomplish, stated precisely
4. Constraints — What NOT to touch or change
5. Expected output — What to report back

Step 3: Dispatch Agents

Use the harness's subagent or task tool to dispatch agents in one parallel operation. If the harness has no subagent capability, keep the work sequential.

Task 1: "Fix SMTP timeout in services/email/..."
Task 2: "Fix auth token refresh in services/auth/..."
Task 3: "Update deploy script error handling in scripts/..."

Choose the right agent type for each task:

Platform built-ins (available in Claude Code regardless of repo config):

• Bash — Command execution, git operations
• general-purpose — Multi-step tasks requiring multiple tools
• Explore — Codebase investigation and search

Repository agents may be available in agents/. Read their metadata before selecting one and do not assume every harness supports the same agent schema.

Step 4: Review Results

After all agents return:

1. Read each summary — Verify the agent accomplished its objective
2. Check for conflicts — Ensure no agent modified files outside its scope
3. Run the full test suite — Confirm no regressions from combined changes
4. Integrate — If agents worked in worktrees, merge branches

Agent Prompt Structure

A well-structured agent prompt follows this template:

You are working on [project context].

## Task
[Precise objective — what to accomplish]

## Scope
- Files to modify: [specific paths]
- Files to read for context: [specific paths]

## Constraints
- Do NOT modify [specific files/areas]
- Do NOT [specific anti-patterns]

## Expected Output
Return:
- [What was changed and why]
- [Any issues encountered]
- [Test results if applicable]

Example: Debugging Multiple BATS Test Files

Task 1 (general-purpose):
"Fix the failing tests in tests/scripts/deploy.bats.
Read scripts/deploy.sh for context. The tests expect
deploy.sh to validate the service name argument before
proceeding. Only modify scripts/deploy.sh — do not
change the test file. Run the tests after fixing and
report results."

Task 2 (general-purpose):
"Fix the failing tests in tests/scripts/secrets.bats.
Read scripts/secrets.sh for context. The tests expect
secrets.sh to check for sops binary before attempting
decryption. Only modify scripts/secrets.sh — do not
change the test file. Run the tests after fixing and
report results."

Example: Researching Independent Topics

Task 1 (Explore):
"Search the codebase for all Traefik configuration.
Find how TLS certificates are managed, what middleware
is configured, and document the routing rules. Return
a summary of the Traefik setup."

Task 2 (Explore):
"Search the codebase for all Keycloak configuration.
Find realm settings, client configurations, and theme
customizations. Return a summary of the auth setup."

Common Mistakes

Mistake	Why It Fails	Fix
Prompt too broad ("fix all the tests")	Agent flounders without focus	Scope to specific files and behaviors
Missing context ("fix the bug")	Agent lacks information to start	Include file paths and background
No constraints ("make it work")	Agent may modify shared files	Specify what NOT to change
Vague output expectation	Can't verify success	Define what the summary must include
Parallelizing dependent tasks	Later agents need earlier results	Identify dependencies first, chain those
Too many agents at once	Diminishing returns, review burden	2-4 agents is the sweet spot

Verification After Agents Return

After all agents complete, always:

1. Review each agent's summary — Did it accomplish the stated objective?
2. Check for scope violations — Did any agent modify files outside its designated area?
3. Look for conflicts — Did agents make contradictory changes?
4. Run the full test suite — Combined changes may interact even if individual tasks were independent
5. Commit atomically or separately — If changes are truly independent, separate commits improve history; if they're part of one feature, a single commit is cleaner

Integration

• using-git-worktrees skill: When agents need filesystem isolation, create a worktree per agent before dispatching.
• using-tmux skill: For agents running in separate terminal sessions, combine worktrees with tmux session-per-agent orchestration.
• Agent patterns: Prefer read-only specialists for research and isolated worktrees for agents that edit files.