| name | multi-agent-validation |
| description | Validate artifacts with multiple independent AI tools — images, code, documents, designs. Use when a single perspective has blind spots, for visual output validation, or when validators disagree. Trigger: multi-agent validation, independent validators, cross-check with another tool, second opinion, validation crew. |
| metadata | {"type":"protocol","invocation":"both","practice":null} |
Multi-Agent Validation
Validate artifacts (images, code, documents, designs) using multiple independent AI tools. Each tool brings different strengths; disagreement surfaces real issues.
When to Use
- Visual output validation (rendered images, UI screenshots, design mockups)
- Code review requiring multiple perspectives (logic, security, performance)
- Documentation quality checks (accuracy, completeness, clarity)
- Design decisions needing independent opinions before committing
- Any artifact requiring qualitative judgment where a single perspective has blind spots
NOT suitable for: purely objective checks (linting, type checking, test pass/fail) where a single authoritative tool suffices.
The Pattern
1. Produce the artifact
2. Run N independent validators (minimum 2, recommended 3)
3. If ANY validator fails → investigate and fix
4. Record all verdicts with evidence
Validator Roles
| Role | Strength | Weakness | Example Tools |
|---|
| Visual Inspector | "Does this look right?" Qualitative impression | Can't measure, may miss technical issues | codex, human eye |
| Technical Auditor | Measures actual values, reads source code | May over-report invisible issues, can misdiagnose root causes | agy, linters, static analysis |
| Contextual Judge | Domain knowledge, intent awareness, breaks ties | Bias toward "knowing what it should do" | kiro, domain expert |
Disagreement Resolution
- All agree PASS → proceed with confidence
- All agree FAIL → fix before proceeding
- Split verdict → investigate the discrepancy:
- Technical auditor wins on measurable criteria (pixel values, code correctness)
- Visual inspector wins on subjective quality (does it look right to a human?)
- Contextual judge breaks ties using domain knowledge and intent
- Auditor finds a bug → verify against authoritative documentation before acting (tools can misdiagnose)
Image Validation (Primary Use Case)
| Tool | Role |
|---|
| kiro (direct image inspection in conversation) | Contextual judge |
codex (codex exec -i <image> --sandbox read-only) | Visual inspector |
agy with --sandbox | Visual inspector + native measurement |
| agy full (reads code) | Technical auditor |
Full command syntax, flags, and per-tool quirks: references/tool-invocation.md.
Prompt Engineering
Critical for toon/NPR/stylized renders — without this context, tools flag intentional stylization as defects:
This is an intentional toon/cel-shaded render where hard shadow edges
are a DESIRED feature (not a bug). PASS if [criteria].
FAIL only if [failure condition].
Structure prompts as checklists for structured, comparable output across tools:
For each [object]:
1. criterion? YES/NO
2. criterion? YES/NO
Overall PASS/FAIL.
Code and Document Validation
Same pattern, different prompts: codex for logic/readability review, agy for security/accuracy audits, kiro for architecture review against project conventions. Command examples: references/lessons-learned.md.
Known Tool Behaviors
Validated characteristics, failure patterns, and token costs live in references/lessons-learned.md. Highlights:
- agy flags intentional styles as defects unless the prompt declares them desired
- agy identifies symptoms correctly but may misattribute root causes — verify against docs
- kiro's standards drift down after many iterations (context fatigue)
Applicability
This pattern works for any domain where quality is partially subjective, no single tool catches all issues, false confidence from one perspective is dangerous, and the cost of shipping a defect exceeds the cost of 2-3 extra checks.
