Supported Frameworks

Agent Lens evaluates any agent that sends traces to MLflow. The framework doesn't matter — only that MLflow receives traces.

Two dimensions of framework independence

1. Agents being evaluated (target agents)

These are the agents whose quality you want to assess. They run independently from Agent Lens and send traces to MLflow.

Framework	MLflow Integration	Autolog
LangGraph	mlflow.langchain.autolog()	One-line — traces flow automatically
Google ADK	mlflow.tracing.enable() or ADK's built-in tracing with MLflow export	Configure MLFLOW_TRACKING_URI on your ADK agent
LangChain	mlflow.langchain.autolog()	One-line — traces flow automatically
CrewAI	mlflow.crewai.autolog()	One-line — traces flow automatically
OpenAI Agents SDK	mlflow.openai.autolog()	One-line — traces flow automatically
AutoGen	mlflow.autogen.autolog() or @mlflow.trace decorator	Autolog or manual
LlamaIndex	mlflow.llama_index.autolog()	One-line — traces flow automatically
Custom Python	@mlflow.trace decorator or usercustomize.py drop-in	See below
Any language	MLflow REST API (POST /api/2.0/mlflow/traces)	HTTP — no SDK needed

2. Agent Lens runtime (qualification harness)

This is the conversational agent that interprets your natural language requests and calls MLflow MCP tools. It ships with Hermes as the reference runtime, but the core artifacts are portable.

Artifact	Portable?	Notes
agent-lens/skills/*.md	Yes	Plain markdown. Any MCP agent can interpret them.
agent-lens/soul.md	Yes	Agent identity and constraints. Standard prompt engineering.
agent-lens/config.yaml	Mostly	MCP URL and tool allowlist are universal.
Containerfile	No	Installs specific harness. Replace for your runtime.
startup.sh	No	Launches harness processes. Replace for your runtime.

Instrumentation examples

LangGraph / LangChain

import mlflow
mlflow.langchain.autolog()

# Your LangGraph agent code runs as usual
# Traces flow to MLflow automatically

Google ADK

import mlflow
mlflow.tracing.enable()

# Or configure ADK's built-in tracing to export to MLflow
# Set MLFLOW_TRACKING_URI in your ADK agent's environment

CrewAI

import mlflow
mlflow.crewai.autolog()

# Your CrewAI agent code runs as usual
# Traces flow to MLflow automatically

OpenAI Agents SDK

import mlflow
mlflow.openai.autolog()

# Your OpenAI agent code runs as usual
# Traces flow to MLflow automatically

AutoGen

import mlflow
mlflow.autogen.autolog()

# Or use the @mlflow.trace decorator on specific functions

Zero-code (any Python agent)

cp instrumentation/usercustomize.py $(python -m site --user-site)/
export MLFLOW_TRACKING_URI="https://your-mlflow:8443"
export MLFLOW_EXPERIMENT_NAME="my-agent"
# No code changes needed — traces flow automatically

Non-Python agents (REST API)

Any agent in any language can send traces via the MLflow REST API:

curl -X POST https://your-mlflow:8443/api/2.0/mlflow/traces \
  -H "Content-Type: application/json" \
  -d '{
    "experiment_id": "1",
    "timestamp_ms": 1721900000000,
    "request": {"messages": [{"role": "user", "content": "Hello"}]},
    "response": {"choices": [{"message": {"content": "Hi there!"}}]}
  }'

What happens after traces arrive

Once traces are in MLflow, Agent Lens can:

1. Observe — search_experiments, search_traces, get_trace
2. Evaluate — evaluate_traces with built-in or custom scorers
3. Annotate — log_trace_feedback, set_trace_tag
4. Qualify — PASS/FAIL verdicts against configurable thresholds
5. Report — Fleet dashboards, executive summaries, compliance exports

The framework that produced the traces is irrelevant. Agent Lens works with the traces, not the agent code.