Agent Dispatch & Structured Prompting

Full agent type definitions: See references/agents.md for complete 37 agent role specifications across 7 swarms (Engineering, Operations, Business, Data, Product, Growth, Review).

How Agents Actually Work

Claude Code's Task tool has these subagent_types:

• general-purpose - Most work (implementation, review, testing)
• Explore - Codebase exploration and search
• Plan - Architecture and planning
• Bash - Command execution
• platform-orchestrator - Deployment and service management

The 37 agent types are ROLES defined through prompts. Create specialized behavior:

# Security reviewer = general-purpose + security-focused prompt
Task(
    subagent_type="general-purpose",
    model="opus",
    description="Security review: auth module",
    prompt="""You are a security reviewer. Focus on:
    - Authentication vulnerabilities
    - Input validation gaps
    - OWASP Top 10 issues
    Review: src/auth/*.ts"""
)

# Frontend agent = general-purpose + frontend-focused prompt
Task(
    subagent_type="general-purpose",
    model="opus",
    description="Implement login form",
    prompt="""You are a frontend developer. Implement:
    - React login form component
    - Form validation
    - Error state handling"""
)

Structured Prompting Template

Every Task dispatch MUST include these sections:

## GOAL
[What success looks like - measurable outcome]

## CONSTRAINTS
[Hard limits - what you cannot do]

## CONTEXT
[Files to read, previous attempts, related decisions]

## OUTPUT
[Exact deliverables expected]

Example:

Task(
    subagent_type="general-purpose",
    model="opus",
    description="Implement user registration API",
    prompt="""
## GOAL
Create POST /api/users endpoint that registers new users.
Success: Endpoint works, tests pass, matches OpenAPI spec.

## CONSTRAINTS
- Use bcrypt for password hashing (already in dependencies)
- No new dependencies without approval
- Response time < 200ms

## CONTEXT
- Existing auth pattern: src/auth/login.ts
- OpenAPI spec: .loki/specs/openapi.yaml
- User model: src/models/user.ts

## OUTPUT
- [ ] Endpoint implementation in src/routes/users.ts
- [ ] Unit tests in tests/users.test.ts
- [ ] Integration test in tests/integration/users.test.ts
"""
)

Parallel Review Pattern

Code review uses 3 parallel reviewers with different focus areas:

# Launch all 3 in ONE message (parallel execution)
Task(
    subagent_type="general-purpose",
    model="opus",
    description="Review: correctness",
    prompt="Review for bugs, logic errors, edge cases. Files: {files}"
)
Task(
    subagent_type="general-purpose",
    model="opus",
    description="Review: security",
    prompt="Review for vulnerabilities, injection, auth issues. Files: {files}"
)
Task(
    subagent_type="general-purpose",
    model="opus",
    description="Review: performance",
    prompt="Review for N+1 queries, memory leaks, slow operations. Files: {files}"
)

Rules:

• ALWAYS use opus for reviews
• ALWAYS launch all 3 in single message
• WAIT for all 3 before aggregating
• IF unanimous approval: run Devil's Advocate reviewer

Confidence-Based Routing

Confidence factors:

• Requirement clarity (30%)
• Similar past successes (20%)
• Technical complexity match (25%)
• Resource availability (15%)
• Time pressure (10%)

Agent Handoffs

When one agent completes and hands off to another:

# Agent A completes, hands off to Agent B
handoff_data = {
    "completed_work": "Implemented user registration endpoint",
    "files_modified": ["src/routes/users.ts", "tests/users.test.ts"],
    "decisions_made": ["Used bcrypt, not argon2", "Email validation via regex"],
    "open_questions": ["Rate limiting not implemented yet"],
    "mistakes_learned": ["First attempt had SQL injection - fixed"]
}

Task(
    subagent_type="general-purpose",
    model="sonnet",
    description="Integration testing for user registration",
    prompt=f"Previous agent completed: {handoff_data}. Now write integration tests..."
)

Project AGENTS.md

IF target project has AGENTS.md, read it first. (OpenAI/AAIF standard)

Priority order for context:

1. AGENTS.md in current directory
2. CLAUDE.md project instructions
3. .loki/CONTINUITY.md session state

A2A-Inspired Communication (Google Protocol)

Agent Cards for capability discovery:

{
  "agent_id": "eng-backend-001",
  "capabilities": ["api-endpoint", "auth", "database"],
  "status": "available",
  "current_task": null,
  "inbox": ".loki/messages/inbox/eng-backend-001/",
  "outbox": ".loki/messages/outbox/eng-backend-001/"
}

Handoff message format:

{
  "from": "eng-backend-001",
  "to": "eng-qa-001",
  "task_id": "task-123",
  "type": "handoff",
  "payload": {
    "completed_work": "POST /api/users implemented",
    "files_modified": ["src/routes/users.ts"],
    "decisions": ["bcrypt for passwords"],
    "artifacts": [".loki/artifacts/users-api-spec.json"]
  }
}

Location: .loki/messages/ directory structure

Agentic Patterns Reference (awesome-agentic-patterns)

Patterns used in Loki Mode:

Key insight (moridinamael.github.io): Simple orchestration beats complex frameworks. "Ralph Wiggum Mode" - basic continuation prompts work better than elaborate coordination systems.

The 37 Agent Roles

See references/agents.md for complete specifications. Summary:

Spawn only what you need. Simple project: 5-10 agents. Complex startup: 100+.