Skip to main content

Generate Data from Your Scenario

Once the manual button test works, let an agent run my_buttons.FirstTask and export the recorded experience. The same preset defines the instruction, available tools, and success condition in both workflows.

Prepare the services​

Complete GEN installation and configuration. Use a compatible model with full-history, or connect your own agent. MAGMA model checkpoints have not been released yet.

Install your scenario package in the Python environment running GEN. If you created it under ~/magma-workspace/my-magma-scenarios:

cd ~/magma-workspace
source .venv/bin/activate
python -m pip install -e ./my-magma-scenarios
magma-scenarios show my_buttons

Adjust the package path if you created it elsewhere. With Conda, activate your MAGMA environment instead of the venv. show should list FirstTask before you continue.

Start the planner, agent, and viewer using the first three terminals in the launch guide. Keep them running. The commands below use the workspace's config.yaml and output/ directory.

Run your task​

In the generation terminal:

cd ~/magma-workspace
source .venv/bin/activate
magma-gen run my_buttons_run --preset my_buttons.FirstTask \
  --config-path ./config.yaml --nb-env 1 --nb-branch 1 \
  --no-coaching --no-judge --no-randomized

Use a new run name each time: the launcher may clear an existing run directory. Compared with the first GEN example, the task selector now points to your installed package.

Select my_buttons_run in the viewer. Inspect the instruction, the agent's proposed tool call, its execution result, and stage completion. A successful button run should call press_button for sw0 and satisfy the button goal. A valid HTTP response alone does not establish task success.

The data is saved in output/my_buttons_run. To use the other button parameter from the tutorial, choose a new run name and append --button sw2 to the command.

Export the result​

For a run produced by full-history:

magma-gen export output/my_buttons_run --agent full-history-agent

Inspect output/my_buttons_run/datasets/ using the export guide. The exporter can return no examples if the graph has no eligible candidates.

For a custom agent, install and select that agent's exporter. The selected exporter must match the agent that produced the graph.

Check answers and multi-stage interactions​

To run the two-button preset, register ButtonSequence and use --preset my_buttons.ButtonSequence with a new run name. Inspect both button actions and the final completion message in the viewer.

For questions, clarification, or rules that require semantic answer checks, first configure a language-model backend. Remove --no-judge from the run command. You can keep --no-coaching while checking your task's answer criteria. Run your registered preset and inspect each answer stage's result, not only the model's text.

If your task uses registered skills, add --skills with the manifest keys to enable them, for example --skills PressPair for the two-button skill. The task must still define goals and boundaries that fit those actions.

Extend the task when needed​

You now have a path from your own task to recorded agent experience. Choose an extension based on your application: