Framework Integration Guide

AgentV is framework-agnostic. You can evaluate agents built with any library (LangChain, AutoGen, CrewAI, etc.) using either Native Adapters or Manual API Wrapping.

🔌 Native Ecosystem Adapters

For the highest level of integration, use native adapters. These typically require a Plugin and allow the harness to communicate directly with the framework’s internal message bus.

Supported Adapters

• langgraph://: Official LangGraph v2 Protocol support.
• crewai://: Support for CrewAI agent swarms.
• ag2://: Support for AG2 (formerly AutoGen) agents.

Usage:

agentv evaluate --agent langgraph://my_retail_node

⚙️ Behavioral Configuration Mesh

As of v1.5.0, native adapters can be granularly configured via the Industrial Confog Mesh (e.g., .aes/config/adapters.d/). This allows you to define behavioral parameters (like Docker usage or custom timeouts) that are mathematically bound to the evaluation environment.

Example: Disabling Docker for AG2 Create .aes/config/adapters.d/ag2_policy.json:

{
  "adapters": {
    "settings": {
      "frameworks": {
        "ag2": {
          "use_docker": false
        }
      }
    }
  }
}

This configuration ensures that all ag2:// evaluations in this environment default to non-Docker execution unless overridden by an environment variable.

---

🛠️ Manual API Wrapping (HTTP/REST)

If a native adapter is not available, you can wrap your agent in a simple REST API that follows the Agent API Contract.

🟢 LangGraph Example

from langgraph.graph import StateGraph
from fastapi import FastAPI

app = FastAPI()


@app.post("/execute_task")
async def execute(request: dict):
    state = {"input": request["input"]}
    result = graph.invoke(state)
    return {"content": result["output"]}

🟠 AG2 Example

import ag2
from fastapi import FastAPI

app = FastAPI()


@app.post("/execute_task")
async def execute(request: dict):
    # Map input to AG2 initiate_chat
    user_proxy.initiate_chat(assistant, message=request["input"])
    return {"content": user_proxy.last_message()["content"]}

🔵 CrewAI Example

from crewai import Crew
from fastapi import FastAPI

app = FastAPI()


@app.post("/execute_task")
async def execute(request: dict):
    crew = Crew(agents=[...], tasks=[...])
    result = crew.kickoff()
    return {"content": result}

---

🚀 Registering Your Agent

Once your agent is running (manually or via a native scheme), register it with the evaluation engine:

agentv evaluate \
  --run-id <id> \
  --agent http://localhost:8000/execute_task \
  --agent-name "Retail-Orchestrator-V1"

📊 Generating Results

After the run completes, generate a Verified Report to analyze the results:

agentv report --run-id <id> --share