# Agent System

## Agent Quick Reference

| Agent | Tier | Model | Purpose |
|-------|------|-------|---------|
| code-critic | 3 | sonnet | Code quality review |
| {domain agents} | {tier} | {model} | {purpose} |

## Consultation Matrix

| Task Type | Mandatory Agent | Recommended Agent |
|-----------|----------------|-------------------|
| {task type} | {agent} | {agent} |

## Design Principles

### Fresh Context for Verification

When verifying work output, spawn a dedicated subagent instead of self-verifying.

**Why**: The producing agent accumulates context bias through planning, decision-making, and execution — it is predisposed to confirm its own output. A fresh subagent has no prior commitment to the result.

**Pattern**:
- Producer agent generates output (files, plans, code)
- Verifier subagent receives only: inputs (requirements, analysis) + outputs (generated files)
- Verifier has a single goal: do the outputs satisfy the inputs?
- One goal, clean context, higher accuracy

**Anti-pattern**: Agent generates artifacts → same agent "spot-checks" its own work → confirmation bias → defects pass through.