Framework guide

Test your AutoGen agent

AutoGen runs multi-agent conversations where agents message one another to solve a task. The failure modes hide in the back-and-forth: an agent that loops, misreads, or never terminates. Connect yours to ClientCoded and every message and turn gets captured and scored.

1. Install

Add the ClientCoded package to your AutoGen project.

Terminal
$ pip install clientcoded

2. Initialize

Add two lines at the top of your AutoGen app, before your agent runs. Your agent_id and team api_key come from your dashboard.

Python
import clientcoded
clientcoded.init(agent_id="your-agent-id", api_key="your-team-api-key")
# Every AutoGen call is now captured and scored.

3. What gets auto-traced

Once init runs, ClientCoded auto-instruments AutoGen. You do not change your agent code. On every run, these are captured automatically:

  • Messages exchanged between conversable agents
  • LLM completions per agent
  • Tool and function calls
  • Group-chat turns and speaker selection
  • Conversation termination and final output

4. See your results

Each run is scored and sent to your dashboard, with the overall grade, the per-dimension breakdown, and the exact step where your agent went wrong.

Output
✓ Connected. Tracing AutoGen calls.

Run scored: 7.8 / 10
15 steps traced · 2 issues flagged
# View the full breakdown in your dashboard

Open your dashboard to see traced runs, scores, and flagged failures. From there you can run the full adversarial test suite against your AutoGen agent and track quality on every change.

Next steps

See the full integration docs for the production monitoring webhook and the Query API, or start free, your first adversarial test is on us.

Testing another framework? LangChain · CrewAI · LlamaIndex