Quick Start¶
This guide walks you through collecting a demonstration, learning a policy, and evaluating the agent.
Prerequisites¶
- OpenAdapt installed with required packages:
pip install openadapt[all] - macOS users: Grant required permissions
1. Collect a Demonstration¶
Start capturing your screen and inputs:
Now perform the task you want to automate:
- Click on applications
- Type text
- Navigate menus
- Complete your workflow
When finished, stop the capture:
2. View the Trajectory¶
Inspect what was captured:
This opens a trajectory viewer showing:
- Observations (screenshots) at each step
- Actions (mouse and keyboard events)
- Timing information
3. List Your Demonstrations¶
See all collected demonstrations:
Output:
4. Learn a Policy¶
Learn an agent policy from your demonstration trajectory:
Monitor policy learning progress:
Policy learning creates a checkpoint file in training_output/.
5. Evaluate the Agent¶
Test your trained policy on a benchmark:
Or run a mock evaluation to verify the setup:
6. Evaluate an API Agent¶
Test API-based agents (Claude, GPT-4V):
# Set your API key
export ANTHROPIC_API_KEY=your-key-here
# Run evaluation
openadapt eval run --agent api-claude --benchmark waa
Complete Workflow Example¶
Here is a complete example demonstrating the full pipeline:
# 1. Install OpenAdapt
pip install openadapt[all]
# 2. Check system requirements
openadapt doctor
# 3. Collect a demonstration
openadapt capture start --name email-reply
# ... perform the task ...
# Press Ctrl+C to stop
# 4. View the trajectory
openadapt capture view email-reply
# 5. Learn a policy
openadapt train start --capture email-reply --model qwen3vl-2b
# 6. Wait for policy learning to complete
openadapt train status
# 7. Evaluate the agent
openadapt eval run --checkpoint training_output/model.pt --benchmark waa
Next Steps¶
- CLI Reference - Full command documentation
- Architecture - How OpenAdapt works
- Packages - Explore individual packages
- Contributing - Help improve OpenAdapt