> ## Documentation Index
> Fetch the complete documentation index at: https://docs-v1.agno.com/llms.txt
> Use this file to discover all available pages before exploring further.

# MLflow

> Instrument Agno agents to send traces, spans, and logs to the MLflow Observability UI.

## Integrating Agno with MLflow Observability

[MLflow](https://mlflow.org) can automatically instrument your Agno agents - capturing agent interactions, model/tool calls, inputs/outputs, and timing and surface them in the **Observability** (Traces) UI.

## Prerequisites

### 1) Install packages

```bash theme={null}
pip install mlflow agno openai
```

> Use your preferred LLM provider package if not OpenAI.

### 2) Tracking server / UI

* **Local**: run an MLflow server/UI or `mlflow ui` for quick local tests.
* **Remote**: point to your team’s MLflow Tracking Server.

### 3) Authentication & configuration

**Standard MLflow env vars:**

```bash theme={null}
export MLFLOW_TRACKING_URI=http://localhost:5000   # or your server
export MLFLOW_TRACKING_TOKEN=<token-if-required>   # only if your server uses tokens
```

## Quickstart: Autologging Agno → MLflow Observability

```python theme={null}
import os
import mlflow
from agno.agent import Agent
from agno.models.openai import OpenAIChat

mlflow.set_tracking_uri(os.getenv("MLFLOW_TRACKING_URI", "http://localhost:5000"))
mlflow.set_experiment("agno-observation-demo")
mlflow.agno.autolog()

agent = Agent(model=OpenAIChat(id="gpt-4o"), name="horror-writer")

with mlflow.start_run(run_name="two-sentence-horror") as run:
    result = agent.run("Share a 2 sentence horror story")
    print(result.content)

    mlflow.log_text(result.content, artifact_file="outputs/horror.txt")

    trace_id = mlflow.get_last_active_trace_id()
    print(f"Trace ID: {trace_id}")
```

### View your traces

Open your MLflow UI and:

* Select the **agno-observation-demo** experiment → open the latest **Run**.
* Go to **Observability / Traces** (or the **Traces** tab in the run) to inspect spans:
  * Agent & model calls (LLM invocations)
  * Tool calls, inputs/outputs
  * Timing, status, and attributes

## Notes

* **Autologging**: Use `mlflow.agno.autolog()` before creating/running agents.
* **Runs & traces**: Wrap your workflow in `with mlflow.start_run(...):` so runs and traces are grouped.
* **Artifacts**: Use helpers like `mlflow.log_text(...)`, `mlflow.log_dict(...)`, or `mlflow.log_figure(...)` to attach outputs to the run.

***

This setup routes Agno’s agent/model/tool activity into MLflow **Observability**, letting you observe every span and correlate rich traces with your experiment runs and artifacts.