Ecosystem Verification

This guide provides the protocols for verifying that your AgentV environment is production-ready across all framework adapters and proprietary models.

🚦 Prerequisites

Environment Setup

Ensure the harness is installed in editable mode and all API keys are configured in your .env file.

pip install -e .

Dependency Audit

Use the doctor utility to perform a global health check of your dependencies and connection status.

agentv doctor

🌎 Framework Verification (Adapters)

Verify that the harness can communicate with external frameworks using the appropriate protocols.

AG2 (formerly AutoGen)

Protocol: ag2

Verification:

agentv run --path scenarios/loan_scenario.json --protocol ag2

LangChain / LangGraph

Protocol: langgraph

Verification:

agentv run --path scenarios/loan_scenario.json --protocol langgraph

CrewAI

Protocol: crewai

Verification:

agentv run --path scenarios/loan_scenario.json --protocol crewai

💎 Proprietary Model Verification

Verify production readiness for frontier models using live API keys.

Provider	Protocol	Verification Command
OpenAI	`openai://`	`agentv run --protocol openai --agent openai://gpt-5.4-mini`
Anthropic	`claude://`	`agentv run --protocol claude --agent claude://claude-4.6-sonnet`
Google	`gemini://`	`agentv run --protocol gemini --agent gemini://gemini-2.5-flash`
xAI	`grok://`	`agentv run --protocol grok --agent grok://grok-4.20-multi-agent`

⚖️ Judge & Calibration Audit

For industrial certification, we must verify that the LLM-Judge is aligned with human ground truth.

1. Rubric Routing

Run a scenario with a specialized rubric and verify the correct prompt injection in the run.jsonl trace.

agentv evaluate --run-id <id>

2. Calibration Command

Compare judge scores against human labels (if present in the trace).

agentv calibrate --run-id <id>

The report will provide Mean Absolute Error (MAE) and Pearson Correlation for the judge.

✅ Production Readiness Checklist

agentv doctor returns all GREEN.
No ImportError on any ecosystem adapter.
API keys are correctly masked in log outputs.
Results generated in reports/ include required framework metadata.
Verification Certificates (VCs) are signed with the project’s private key.