Agent Coordination Discipline

Agent Coordination Discipline

Iron Law: "NO AGENT LAUNCH WITHOUT CLEAR DELEGATION CRITERIA"

When to Use

Use this skill when:

• Considering launching an agent with the Task tool
• Evaluating whether a task requires agent delegation
• Selecting between different agent types or external models
• Coordinating multiple agents in a workflow
• Implementing external-model for external model delegation
• Debugging agent coordination failures

This skill prevents premature agent launches, redundant agent usage, and poor task isolation that wastes thinking budget and causes coordination failures.

Red Flags (Violation Indicators)

• Agent for single grep - Launching agent to run one grep/glob command (trivial-task anti-pattern)
• Missing external-model model - Using external-model without explicit model name specification
• No task isolation - Agent task description lacks independent context or success criteria
• No success criteria - Task description doesn't define what "done" looks like
• Default thinking pattern - Not considering whether task needs deep thinking vs. fast execution
• Multiple agents without coordination - Launching 2+ agents without clear result routing plan
• Result not used - Launching agent but not routing/validating its output
• Agent for trivial decision - Using agent to make decision you could make directly
• No tool exhaustion check - Launching agent before trying native tools first
• Missing timeout consideration - Not evaluating if task needs extended thinking time
• No error handling plan - Not defining what happens if agent fails or returns partial results
• Skill gap unclear - Not identifying what specific expertise the agent provides

Key Concepts

1. Agent vs. Native Tools Decision Tree

Does the task require:
├─ Single tool call (grep, read, edit)?
│  └─ ✗ NO AGENT - Use native tool directly
├─ 2-3 sequential tool calls?
│  └─ ✗ NO AGENT - Use tools directly in sequence
├─ Multi-step investigation with branching logic?
│  └─ ✓ AGENT - Task tool with developer/architect agent
├─ External model expertise (Grok, DeepSeek, etc.)?
│  └─ ✓ AGENT - external-model pattern with model specification
├─ Parallel exploration of multiple code paths?
│  └─ ✓ AGENT - Multiple Task calls with coordination
└─ High-risk change needing isolation?
   └─ ✓ AGENT - Task tool with sandbox/review focus

2. Task Isolation Requirements

Every agent task must be independently executable:

Bad (not isolated):

Task: "Fix the bug we discussed earlier"

Good (properly isolated):

Task: "Debug the TypeError in src/components/UserProfile.tsx line 42.
Context: User reports 'Cannot read property name of undefined' when viewing profile page.
Evidence: Error occurs after recent commit abc123 that changed user data structure.
Success criteria: Identify root cause, propose fix, verify with test scenario."

3. External Model Pattern

When delegating to external models via claudish CLI:

Structure:

claudish --model {model_id} --stdin --quiet <<EOF > output.md
{Task Description}

Context:
- {Relevant file paths}
- {Current state}
- {Related decisions}

Success Criteria:
- {What constitutes success}
- {Expected output format}

Constraints:
- {Time limits}
- {Tool restrictions}
- {Quality requirements}
EOF

Example:

claudish --model x-ai/grok-code-fast-1 --stdin --quiet <<EOF > analysis.md
Analyze the React component rendering performance issue in Dashboard.tsx.

Context:
- File: src/components/Dashboard.tsx (247 lines)
- Issue: Component re-renders 40+ times on data updates
- Recent changes: Added real-time WebSocket updates in commit f4a2c1b

Success Criteria:
- Identify unnecessary re-renders (provide line numbers)
- Propose memoization strategy
- Estimate performance improvement

Constraints:
- Max 3 minutes analysis time
- Focus on React 19 compiler-friendly patterns
EOF

When to Use Agents

Multi-Step Investigation

Trigger: Task requires 5+ tool calls with conditional branching Agent: developer, architect Example: "Trace data flow through 3 layers to find where user.email becomes null"

External Model Expertise

Trigger: Need specialized model capabilities (code speed, vision, reasoning) Agent: external-model with specific model Example: "Use Grok Code Fast to refactor 15 files for consistency in < 2 minutes"

Parallel Work

Trigger: Multiple independent tasks that can run simultaneously Agent: Multiple Task calls with result aggregation Example: "Analyze frontend performance (Task 1) while auditing API security (Task 2)"

Risk Isolation

Trigger: High-risk changes needing review before merging to main workflow Agent: review-focused agent with checkpoint Example: "Evaluate if this database migration will cause downtime"

Skill Gaps

Trigger: Current agent lacks specific skill that another agent has Agent: specialist agent (security, performance, accessibility) Example: "Launch accessibility agent to audit ARIA compliance"

When NOT to Use Agents

Single Grep/Glob

Instead: Use native Grep or Glob tool directly

# ✗ DON'T
Task: "Find all files using the deprecated API"

# ✓ DO
Grep("oldApiCall", output_mode: "files_with_matches", type: "js")

Simple Tool Execution

Instead: Use tool directly

# ✗ DON'T
Task: "Read the config file and tell me the API URL"

# ✓ DO
Read("/path/to/config.json")
// Parse and extract apiUrl field

Decision Already Made

Instead: Execute the decision

# ✗ DON'T
Task: "I think we should use React Query. What do you think?"

# ✓ DO
// Just implement React Query since decision is made
Write("src/hooks/useApiQuery.ts", reactQueryCode)

Sequential Tool Calls

Instead: Chain tools directly

# ✗ DON'T
Task: "Find the function, read it, and edit it"

# ✓ DO
Grep("functionName", output_mode: "files_with_matches")
// => result: src/utils/helper.ts
Read("src/utils/helper.ts")
Edit("src/utils/helper.ts", old_string, new_string)

Nuanced Context Required

Instead: Handle in current agent

# ✗ DON'T
Task: "Based on our earlier discussion about performance vs. maintainability trade-offs, decide if we should cache this"

# ✓ DO
// Current agent already has context, make decision directly
if (performanceIsCritical) {
  implementCaching()
}

Agent Selection Matrix

external-model Pattern Details

1. Model Selection

Fast Execution (< 2 min):

• x-ai/grok-code-fast-1 - Code generation, refactoring, simple analysis
• anthropic/claude-3-5-haiku - Quick decisions, data transformation

Deep Reasoning (> 2 min):

• anthropic/claude-sonnet-4-5 - Complex debugging, architecture design
• google/gemini-2.0-flash-thinking-exp-01-21 - Extended thinking budget

Specialized:

• Vision models - Screenshot analysis, diagram interpretation
• Code models - Language-specific optimization

2. Context Packaging

Minimal (< 1000 tokens):

• File paths only
• Error message
• Success criteria

Moderate (1000-5000 tokens):

• Key code snippets (< 50 lines)
• Related file structure
• Recent commit context

Full (5000+ tokens):

• Complete file contents
• Related test files
• Architecture documentation