🔬

CMVK Review

Cross-Model Verification Kit: Multi-model consensus for agent safety, correctness, and policy compliance.

What is CMVK?

CMVK (Cross-Model Verification Kit) is AgentOS's unique approach to ensuring agent safety. Instead of relying on a single AI model's judgment, CMVK submits your agent code to multiple independent AI models for review. Only when all models reach consensus is the agent approved for deployment.

Consensus Process

GPT-4 Claude 3 Gemini Pro

↓ ↓ ↓

✅ Safe ✅ Safe ✅ Safe

✅ Consensus Reached: Approved

Participating Models

🤖

GPT-4

OpenAI's flagship model

🧠

Claude 3

Anthropic's safety-focused model

💎

Gemini Pro

Google's advanced model

🦙

Llama 3

Meta's open model

Running CMVK Review

1 Request Review

User: @agentos review

AgentOS: 🔬 Initiating CMVK Review

Submitting to verification models...
├── GPT-4: Analyzing...
├── Claude 3: Analyzing...
├── Gemini Pro: Analyzing...
└── Llama 3: Analyzing...

Please wait while models review your agent.
This typically takes 10-30 seconds.

2 Review Results

AgentOS: 🔬 CMVK Review Complete

Consensus: ✅ APPROVED (4/4 models agree)

Individual Verdicts:

✅ GPT-4: Approved

Agent follows best practices for data handling. No security concerns detected. Rate limiting is properly implemented.

✅ Claude 3: Approved

Safety policies are comprehensive. Error handling prevents data leaks. Logging excludes PII as required.

✅ Gemini Pro: Approved

Code structure is clean and maintainable. Resource usage is within acceptable limits.

✅ Llama 3: Approved

No unsafe operations detected. API calls are properly authenticated.

Handling Rejections

When models don't reach consensus, you'll see detailed feedback:

AgentOS: 🔬 CMVK Review Complete

Consensus: ❌ NOT REACHED (2/4 models raised concerns)

Individual Verdicts:

✅ GPT-4: Approved

Core functionality is safe and well-implemented.

❌ Claude 3: Rejected

Concern: Line 45 logs user email addresses without masking. This violates GDPR data minimization requirements.

⚠️ Gemini Pro: Concerns

Concern: No rate limiting on external API calls. Could lead to quota exhaustion or abuse.

✅ Llama 3: Approved

No critical issues found.

Required Actions:
1. Mask email addresses in log statements (line 45)
2. Add rate limiting for external API calls

Fix these issues and run @agentos review again.

Review Categories

CMVK evaluates agents across multiple dimensions:

🔒 Security

Credentials handling, injection vulnerabilities, authentication

🛡️ Safety

Policy compliance, action boundaries, failure modes

📋 Compliance

GDPR, HIPAA, SOC2, PCI-DSS requirements

⚡ Reliability

Error handling, retry logic, graceful degradation

📊 Resources

Memory usage, API quotas, execution time

Configuration

Customize CMVK review settings:

User: @agentos review --strict

# Strict mode requires 100% consensus
# Default mode allows 75% consensus (3/4 models)

Review Modes:
--strict     All models must approve
--standard   75% consensus required (default)
--quick      Uses 2 models for faster review

💡 Recommendation: Use --strict mode for production deployments and agents handling sensitive data. Use --quick mode during development for faster iteration.

Why Multi-Model?

Different AI models have different strengths and blind spots. By requiring consensus across multiple models from different providers, CMVK catches issues that any single model might miss:

Claude excels at safety and ethical considerations
GPT-4 is strong at code quality and best practices
Gemini catches performance and resource issues
Llama provides an open-source perspective

⚠️ Important: CMVK review is required before deploying agents to production. Agents that bypass CMVK review will be automatically blocked by the AgentOS runtime.

← GitHub Deployment Back to Overview →