launch-sub-agent

Decision Router

IF task is trivial lookup or simple transformation → Haiku (cost-effective) IF task requires architecture, novel problem-solving, or critical decisions → Opus (maximum reasoning) IF task is non-complex but output is extensive → Sonnet (balanced capability and cost) IF task matches specialized domain (code, research, architecture) → Select matching agent type IF default/uncertain → Opus (quality over cost)

launch-sub-agent

Launch a focused sub-agent with intelligent model selection and mandatory verification.

Policy: Why This Pattern

Context isolation is the primary benefit. Each sub-agent operates in a clean context window focused on its specific task. Evaluating work within the same context that produced it creates confirmation bias.

The pattern has three mandatory components:

1. Zero-shot Chain-of-Thought reasoning prefix (think before acting)
2. Task body with scope and constraints
3. Self-critique verification suffix (verify before completing)

Mechanism: Five-Phase Dispatch

Phase 1: Task Analysis

Analyze systematically before dispatching:

1. Task Type: Code implementation / Research / Documentation / Review / Architecture / Testing / Simple transformation
2. Complexity: High (architecture, novel problems) / Medium (standard patterns) / Low (simple, well-defined)
3. Output Size: Large (multiple files) / Medium (single feature) / Small (quick answer)
4. Domain Match: Does task align with specialized agent profile?

Phase 2: Model Selection

Task Profile	Model	Rationale
Complex reasoning (architecture, design, critical)	Opus	Maximum capability
Specialized domain + complex	Opus + agent	Expertise + reasoning
Simple and short	Haiku	Fast, cost-effective
Long output but not complex	Sonnet	Balanced
Default/uncertain	Opus	Quality over cost

Decision tree:

Is task COMPLEX?
├── YES → Opus
│         └── Match specialized domain? → Include agent prompt
└── NO → Is task SIMPLE + SHORT?
          ├── YES → Haiku
          └── NO → Sonnet (long output, not complex)
                  └── Otherwise → Opus (default)

Phase 3: Specialized Agent Matching

When task matches a specialized domain, incorporate relevant agent instructions after the CoT prefix.

Use specialized agents when:

• Domain expertise clearly improves quality
• Task clearly fits a known pattern (developer, researcher, architect)

Skip when:

• Task is trivial (specialization overhead not justified)
• No matching agent available

Phase 4: Construct Prompt

Build with three mandatory sections:

CoT Prefix (first):

## Reasoning Approach

Before taking any action, think systematically:

1. What is the core objective? What are the explicit requirements?
2. What are the major components? What order should I tackle them?
3. What assumptions am I making? What edge cases might exist?
4. Does my approach address all requirements? Is there a simpler way?

Task Body:

<task>{task description}</task>
<constraints>{any constraints from context}</constraints>
<context>{relevant files, patterns, requirements}</context>
<output>{expected deliverable format and location}</output>

Self-Critique Suffix (last):

## Self-Critique (MANDATORY)

Before completing, verify your work:

1. Generate 5 verification questions specific to this task
2. Answer each with specific evidence from your solution
3. If ANY gap found: STOP, FIX, RE-VERIFY, DOCUMENT

Do not submit until all verification questions have satisfactory answers.

Phase 5: Dispatch

Use Task tool with selected model. Pass only context relevant to this task (not entire conversation history).

Examples

Input	Analysis	Selection
"Design caching strategy for 10k req/s"	Architecture, High complexity	Opus + architect agent
"Update README with --verbose flag"	Documentation, simple	Haiku
"Implement pagination for /users"	Code, medium complexity	Sonnet + developer agent
"Research auth options (OAuth2, SAML, passwordless)"	Research, high complexity	Opus + researcher agent

Key Principles

• Self-critique loop is non-negotiable
• Sub-agents must answer verification questions before completing
• Context isolation: pass only relevant context, not conversation history
• When in doubt, use Opus (quality over cost)