dispatching-parallel-agents

name	dispatching-parallel-agents
description	Use when facing 3+ independent failures that can be investigated without shared state or dependencies - dispatches multiple agents to investigate and fix independent problems concurrently

<kimi_compat> This skill was ported to Kimi Code CLI. In Kimi:

• Load skills with /skill:name or let the model invoke them automatically.
• "SetTodoList" means use the SetTodoList tool.
• "Agent()" / "Agent" means use the built-in Agent tool.
• "parallel Agent calls with run_in_background=true" means dispatch multiple Agent calls with run_in_background=true.
• Claude-specific hooks and slash commands are not available; skip those steps. </kimi_compat> <skill_overview> When facing 3+ independent failures, dispatch one agent per problem domain to investigate concurrently; verify independence first, dispatch all in single message, wait for all agents, check conflicts, verify integration. </skill_overview>

<rigidity_level> MEDIUM FREEDOM - Follow the 6-step process (identify, create tasks, dispatch, monitor, review, verify) strictly. Independence verification mandatory. Parallel dispatch in single message required. Adapt agent prompt content to problem domain. </rigidity_level>

<quick_reference>

Step	Action	Critical Rule
1. Identify Domains	Test independence (fix A doesn't affect B)	3+ independent domains required
2. Create Agent Tasks	Write focused prompts (scope, goal, constraints, output)	One prompt per domain
3. Dispatch Agents	Launch all agents in SINGLE message	Multiple Agent calls in parallel
4. Monitor Progress	Track completions, don't integrate until ALL done	Wait for all agents
5. Review Results	Read summaries, check conflicts	Manual conflict resolution
6. Verify Integration	Run full test suite	Use verification-before-completion

Why 3+? With only 2 failures, coordination overhead often exceeds sequential time.

Critical: Dispatch all agents in a single parallel tool call, or they run sequentially. </quick_reference>

<when_to_use> Use when:

• 3+ test files failing with different root causes
• Multiple subsystems broken independently
• Each problem can be understood without context from others
• No shared state between investigations
• You've verified failures are truly independent
• Each domain has clear boundaries (different files, modules, features)

Don't use when:

• Failures are related (fix one might fix others)
• Need to understand full system state first
• Agents would interfere (editing same files)
• Haven't verified independence yet (exploratory phase)
• Failures share root cause (one bug, multiple symptoms)
• Need to preserve investigation order (cascading failures)
• Only 2 failures (overhead exceeds benefit) </when_to_use>

<the_process>

Step 1: Identify Independent Domains

Announce: "I'm using /skill:dispatching-parallel-agents to investigate these independent failures concurrently."

Create checklist tracker:

- Identify independent domains (3+ domains identified)
- Create agent tasks (one prompt per domain drafted)
- Dispatch agents in parallel (all agents launched in single message)
- Monitor agent progress (track completions)
- Review results (summaries read, conflicts checked)
- Verify integration (full test suite green)

Test for independence:

1. Ask: "If I fix failure A, does it affect failure B?"

   • If NO → Independent
   • If YES → Related, investigate together

2. Check: "Do failures touch same code/files?"

   • If NO → Likely independent
   • If YES → Check if different functions/areas

3. Verify: "Do failures share error patterns?"

   • If NO → Independent
   • If YES → Might be same root cause

Example independence check:

Failure 1: Authentication tests failing (auth.test.ts)
Failure 2: Database query tests failing (db.test.ts)
Failure 3: API endpoint tests failing (api.test.ts)

Check: Does fixing auth affect db queries? NO
Check: Does fixing db affect API? YES - API uses db

Result: 2 independent domains:
  Domain 1: Authentication (auth.test.ts)
  Domain 2: Database + API (db.test.ts + api.test.ts together)

Group failures by what's broken:

• File A tests: Tool approval flow
• File B tests: Batch completion behavior
• File C tests: Abort functionality

---

Step 2: Create Focused Agent Tasks

Each agent prompt must have:

1. Specific scope: One test file or subsystem
2. Clear goal: Make these tests pass
3. Constraints: Don't change other code
4. Expected output: Summary of what you found and fixed

Good agent prompt example:

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 by:
   - Replacing arbitrary timeouts with event-based waiting
   - Fixing bugs in abort implementation if found
   - Adjusting test expectations if testing changed behavior

Never just increase timeouts - find the real issue.

Return: Summary of what you found and what you fixed.

What makes this good:

• Specific test failures listed
• Context provided (timing/race conditions)
• Clear methodology (read, identify, fix)
• Constraints (don't just increase timeouts)
• Output format (summary)

Common mistakes:

❌ Too broad: "Fix all the tests" - agent gets lost ✅ Specific: "Fix agent-tool-abort.test.ts" - focused scope

❌ No context: "Fix the race condition" - agent doesn't know where ✅ Context: Paste the error messages and test names

❌ No constraints: Agent might refactor everything ✅ Constraints: "Do NOT change production code" or "Fix tests only"

❌ Vague output: "Fix it" - you don't know what changed ✅ Specific: "Return summary of root cause and changes"

---

Step 3: Dispatch All Agents in Parallel

CRITICAL: Dispatch all agents in a SINGLE parallel tool call. Use Agent inside parallel Agent calls with run_in_background=true so they run concurrently.

{
  "tool": "parallel Agent calls with run_in_background=true",
  "tool_uses": [
    {
      "recipient_name": "kimi_cli.tools.agent:Agent",
      "parameters": {
        "subagent_type": "coder",
        "message": "Fix agent-tool-abort.test.ts failures. Return summary of root cause and changes."
      }
    },
    {
      "recipient_name": "kimi_cli.tools.agent:Agent",
      "parameters": {
        "subagent_type": "coder",
        "message": "Fix batch-completion-behavior.test.ts failures. Return summary of root cause and changes."
      }
    },
    {
      "recipient_name": "kimi_cli.tools.agent:Agent",
      "parameters": {
        "subagent_type": "coder",
        "message": "Fix tool-approval-race-conditions.test.ts failures. Return summary of root cause and changes."
      }
    }
  ]
}

Avoid: sequentially calling Agent in separate messages. That runs agents one at a time.

After dispatch:

• Mark "Dispatch agents in parallel" as completed in the checklist
• Mark "Monitor agent progress" as in_progress
• Wait for all agents to complete before integration

---

Step 4: Monitor Progress

As agents work:

• Note which agents have completed
• Note which are still running
• Don't start integration until ALL agents done

If an agent gets stuck (>5 minutes):

1. Check AgentOutput to see what it's doing
2. If stuck on wrong path: Cancel and retry with clearer prompt
3. If needs context from other domain: Wait for other agent, then restart with context
4. If hit real blocker: Investigate blocker yourself, then retry

---

Step 5: Review Results and Check Conflicts

When all agents return:

1. Read each summary carefully

   • What was the root cause?
   • What did the agent change?
   • Were there any uncertainties?

2. Check for conflicts

   • Did multiple agents edit same files?
   • Did agents make contradictory assumptions?
   • Are there integration points between domains?

3. Integration strategy:

   • If no conflicts: Apply all changes
   • If conflicts: Resolve manually before applying
   • If assumptions conflict: Verify with user

4. Document what happened

   • Which agents fixed what
   • Any conflicts found
   • Integration decisions made

---

Step 6: Verify Integration

Run full test suite:

• Not just the fixed tests
• Verify no regressions in other areas
• Use /skill:verification-before-completion skill

Before completing:

# Run all tests
npm test  # or cargo test, pytest, etc.

# Verify output
# If all pass → Mark "Verify integration" complete
# If failures → Identify which agent's change caused regression

</the_process>

Developer dispatches agents sequentially instead of in parallel # Developer sees 3 independent failures # Creates 3 agent prompts

Dispatches first agent

Agent(prompt1)

Waits for response from agent 1

Then dispatches second agent

Agent(prompt2)

Waits for response from agent 2

Then dispatches third agent

Agent(prompt3)

Total time: Sum of all three agents (sequential)

<why_it_fails>

• Agents run sequentially, not in parallel
• No time savings from parallelization
• Each agent waits for previous to complete
• Defeats entire purpose of parallel dispatch
• Same result as sequential investigation
• Wasted overhead of creating separate agents </why_it_fails>

**Dispatch all agents in SINGLE message:**

{
  "tool": "parallel Agent calls with run_in_background=true",
  "tool_uses": [
    {
      "recipient_name": "kimi_cli.tools.agent:Agent",
      "parameters": {
        "subagent_type": "coder",
        "message": "Fix the 3 failing tests in src/agents/agent-tool-abort.test.ts: [prompt 1 content]"
      }
    },
    {
      "recipient_name": "kimi_cli.tools.agent:Agent",
      "parameters": {
        "subagent_type": "coder",
        "message": "Fix the 2 failing tests in src/agents/batch-completion-behavior.test.ts: [prompt 2 content]"
      }
    },
    {
      "recipient_name": "kimi_cli.tools.agent:Agent",
      "parameters": {
        "subagent_type": "coder",
        "message": "Fix the 1 failing test in src/agents/tool-approval-race-conditions.test.ts: [prompt 3 content]"
      }
    }
  ]
}

What happens:

• All three agents start simultaneously
• Each investigates independently
• All complete in parallel
• Total time: Max(agent1, agent2, agent3) instead of Sum

What you gain:

• True parallelization - 3 problems solved concurrently
• Time saved: 3 investigations in time of 1
• Each agent focused on narrow scope
• No waiting for sequential completion
• Proper use of parallel dispatch pattern

Developer assumes failures are independent without verification # Developer sees 3 test failures: # - API endpoint tests failing # - Database query tests failing # - Cache invalidation tests failing

Thinks: "Different subsystems, must be independent"

Dispatches 3 agents immediately without checking independence

Agent 1 finds: API failing because database schema changed

Agent 2 finds: Database queries need migration

Agent 3 finds: Cache keys based on old schema

All three failures caused by same root cause: schema change

Agents make conflicting fixes based on different assumptions

Integration fails because fixes contradict each other

<why_it_fails>

• Skipped independence verification (Step 1)
• Assumed independence based on surface appearance
• All failures actually shared root cause (schema change)
• Agents worked in isolation without seeing connection
• Each agent made different assumptions about correct schema
• Conflicting fixes can't be integrated
• Wasted time on parallel work that should have been unified
• Have to throw away agent work and start over </why_it_fails>

**Run independence check FIRST:**

Check: Does fixing API affect database queries?
- API uses database
- If database schema changes, API breaks
- YES - these are related

Check: Does fixing database affect cache?
- Cache stores database results
- If database schema changes, cache keys break
- YES - these are related

Check: Do failures share error patterns?
- All mention "column not found: user_email"
- All started after schema migration
- YES - shared root cause

Result: NOT INDEPENDENT
These are one problem (schema change) manifesting in 3 places

Correct approach:

Single agent investigates: "Schema migration broke 3 subsystems"

Agent prompt:
"We have 3 test failures all related to schema change:
1. API endpoints: column not found
2. Database queries: column not found
3. Cache invalidation: old keys

Investigate the schema migration that caused this.
Fix by updating all 3 subsystems consistently.
Return: What changed in schema, how you fixed each subsystem."

# One agent sees full picture
# Makes consistent fix across all 3 areas
# No conflicts, proper integration

What you gain:

• Caught shared root cause before wasting time
• One agent sees full context
• Consistent fix across all affected areas
• No conflicting assumptions
• No integration conflicts
• Faster than 3 agents working at cross-purposes
• Proper problem diagnosis before parallel dispatch

Developer integrates agent results without checking conflicts # 3 agents complete successfully # Developer quickly reads summaries:

Agent 1: "Fixed timeout issue by increasing wait time to 5000ms" Agent 2: "Fixed race condition by adding mutex lock" Agent 3: "Fixed timing issue by reducing wait time to 1000ms"

Developer thinks: "All agents succeeded, ship it"

Applies all changes without checking conflicts

Result:

- Agent 1 and Agent 3 edited same file

- Agent 1 increased timeout, Agent 3 decreased it

- Final code has inconsistent timeouts

- Agent 2's mutex interacts badly with Agent 3's reduced timeout

- Tests still fail after integration

<why_it_fails>

• Skipped conflict checking (Step 5)
• Didn't carefully read what each agent changed
• Agents made contradictory decisions
• Agent 1 and Agent 3 had different assumptions about timing
• Agent 2's locking interacts with timing changes
• Blindly applying all fixes creates inconsistent state
• Tests fail after "successful" integration
• Have to manually untangle conflicting changes </why_it_fails>

**Review results carefully before integration:**

## Agent Summaries Review

Agent 1: Fixed timeout issue by increasing wait time to 5000ms
- File: src/agents/tool-executor.ts
- Change: DEFAULT_TIMEOUT = 5000

Agent 2: Fixed race condition by adding mutex lock
- File: src/agents/tool-executor.ts
- Change: Added mutex around tool execution

Agent 3: Fixed timing issue by reducing wait time to 1000ms
- File: src/agents/tool-executor.ts
- Change: DEFAULT_TIMEOUT = 1000

## Conflict Analysis

**CONFLICT DETECTED:**
- Agents 1 and 3 edited same file (tool-executor.ts)
- Agents 1 and 3 changed same constant (DEFAULT_TIMEOUT)
- Agent 1: increase to 5000ms
- Agent 3: decrease to 1000ms
- Contradictory assumptions about correct timing

**Why conflict occurred:**
- Domains weren't actually independent (same timeout constant)
- Both agents tested locally, didn't see interaction
- Different problem spaces led to different timing needs

## Resolution

**Option 1:** Different timeouts for different operations
```typescript
const TOOL_EXECUTION_TIMEOUT = 5000  // Agent 1's need
const TOOL_APPROVAL_TIMEOUT = 1000   // Agent 3's need

Option 2: Investigate why timing varies

• Maybe Agent 1's tests are actually slow (fix slowness)
• Maybe Agent 3's tests are correct (use 1000ms everywhere)

Choose Option 2 after investigation:

• Agent 1's tests were slow due to unrelated issue
• Fix the slowness, use 1000ms timeout everywhere
• Agent 2's mutex is compatible with 1000ms

Integration steps:

1. Apply Agent 2's mutex (no conflict)
2. Apply Agent 3's 1000ms timeout
3. Fix Agent 1's slow tests (root cause)
4. Don't apply Agent 1's timeout increase (symptom fix)

**Run full test suite:**
```bash
npm test
# All tests pass ✅

What you gain:

• Caught contradiction before breaking integration
• Understood why agents made different decisions
• Resolved conflict thoughtfully, not arbitrarily
• Fixed root cause (slow tests) not symptom (long timeout)
• Verified integration works correctly
• Avoided shipping inconsistent code
• Professional conflict resolution process

<failure_modes>

Agent Gets Stuck

Symptoms: No progress after 5+ minutes

Causes:

• Prompt too vague, agent exploring aimlessly
• Domain not actually independent, needs context from other agents
• Agent hit a blocker (missing file, unclear error)

Recovery:

1. Use AgentOutput tool to check what it's doing
2. If stuck on wrong path: Cancel and retry with clearer prompt
3. If needs context from other domain: Wait for other agent, then restart with context
4. If hit real blocker: Investigate blocker yourself, then retry

---

Agents Return Conflicting Fixes

Symptoms: Agents edited same code differently, or made contradictory assumptions

Causes:

• Domains weren't actually independent
• Shared code between domains
• Agents made different assumptions about correct behavior

Recovery:

1. Don't apply either fix automatically
2. Read both fixes carefully
3. Identify the conflict point
4. Resolve manually based on which assumption is correct
5. Consider if domains should be merged

---

Integration Breaks Other Tests

Symptoms: Fixed tests pass, but other tests now fail

Causes:

• Agent changed shared code
• Agent's fix was too broad
• Agent misunderstood requirements

Recovery:

1. Identify which agent's change caused the regression
2. Read the agent's summary - did they mention this change?
3. Evaluate if change is correct but tests need updating
4. Or if change broke something, need to refine the fix
5. Use /skill:verification-before-completion skill for final check

---

False Independence

Symptoms: Fixing one domain revealed it affected another

Recovery:

1. Merge the domains
2. Have one agent investigate both together
3. Learn: Better independence test needed upfront </failure_modes>

<critical_rules>

Rules That Have No Exceptions

1. Verify independence first → Test with questions before dispatching
2. 3+ domains required → 2 failures: overhead exceeds benefit, do sequentially
3. Single message dispatch → All agents in one parallel tool call with multiple Agent uses
4. Wait for ALL agents → Don't integrate until all complete
5. Check conflicts manually → Read summaries, verify no contradictions
6. Verify integration → Run full suite yourself, don't trust agents
7. Checklist tracking → Track agent progress explicitly

Common Excuses

All of these mean: STOP. Follow the process.

• "Just 2 failures, can still parallelize" (Overhead exceeds benefit, do sequentially)
• "Probably independent, will dispatch and see" (Verify independence FIRST)
• "Can dispatch sequentially to save syntax" (WRONG - must dispatch in single message)
• "Agent failed, but others succeeded - ship it" (All agents must succeed or re-investigate)
• "Conflicts are minor, can ignore" (Resolve all conflicts explicitly)
• "Don't need a checklist for just tracking agents" (Use a checklist, track properly)
• "Can skip verification, agents ran tests" (Agents can make mistakes, YOU verify) </critical_rules>

<verification_checklist> Before completing parallel agent work:

• Verified independence with 3 questions (fix A affects B? same code? same error pattern?)
• 3+ independent domains identified (not 2 or fewer)
• Created focused agent prompts (scope, goal, constraints, output)
• Dispatched all agents in a single parallel tool call (multiple Agent uses)
• Waited for ALL agents to complete (didn't integrate early)
• Read all agent summaries carefully
• Checked for conflicts (same files, contradictory assumptions)
• Resolved any conflicts manually before integration
• Ran full test suite (not just fixed tests)
• Used verification-before-completion skill
• Documented which agents fixed what

Can't check all boxes? Return to the process and complete missing steps. </verification_checklist>

**This skill covers:** Parallel investigation of independent failures

Related skills:

• /skill:debugging-with-tools (how to investigate individual failures)
• /skill:fixing-bugs (complete bug workflow)
• /skill:verification-before-completion (verify integration)
• /skill:test-runner (run tests without context pollution)

This skill uses:

• Task tool (dispatch parallel agents)
• AgentOutput tool (monitor stuck agents)
• Checklist (track agent progress)

Workflow integration:

Multiple independent failures
    ↓
Verify independence (Step 1)
    ↓
Create agent tasks (Step 2)
    ↓
Dispatch in parallel (Step 3)
    ↓
Monitor progress (Step 4)
    ↓
Review + check conflicts (Step 5)
    ↓
Verify integration (Step 6)
    ↓
/skill:verification-before-completion

Real example from session (2025-10-03):

• 6 failures across 3 files
• 3 agents dispatched in parallel
• All investigations completed concurrently
• All fixes integrated successfully
• Zero conflicts between agent changes
• Time saved: 3 problems solved in parallel vs sequentially

**Key principles:** - Parallelization only wins with 3+ independent problems - Independence verification prevents wasted parallel work - Single message dispatch is critical for true parallelism - Conflict checking prevents integration disasters - Full verification catches agent mistakes

When stuck:

• Agent not making progress → Check AgentOutput, retry with clearer prompt
• Conflicts after dispatch → Domains weren't independent, merge and retry
• Integration fails tests → Identify which agent caused regression
• Unclear if independent → Test with 3 questions (affects? same code? same error?)