# AutoGen

Instrument AutoGen conversable agents with a one-line patch.

---

AutoGen instruments through a single patch call. It wraps `ConversableAgent`, so each agent's LLM turn in a conversation becomes a TokenJam span without changes to your agent setup.

## Setup

```python
from tokenjam.sdk.integrations.autogen import patch_autogen

patch_autogen()
```

Call it once at startup, before you build your agents or start a conversation.

## What gets captured

Every turn a `ConversableAgent` takes emits a span with token usage, timing, and the model behind it. In a multi-agent conversation you get one trace covering the exchange, with each agent's turns as spans you can attribute cost to individually.

The patch is framework-agnostic and composes with the provider patches. When an AutoGen agent reaches an API client tj already patches, inner spans are de-duplicated so nothing is double-counted. See the [Frameworks overview](/docs/frameworks) for the full picture.