Tutorials

These paths connect setup, execution, inspection, and the next decision. Choose the path that matches your intended outcome and the environment you control.

Complete your first Desktop task

1. Install Desktop, choose a local repository, and confirm which model and tool connections are enabled.
2. Start with a bounded question in Chat. Switch to Work when the task requires files, commands, or several steps; see Chat and Work.
3. Ask Pal for a concrete outcome and state any constraints. Review requested access and approvals in context.
4. Inspect the resulting files and changes in the project. Verify the task outcome before accepting or sharing it.
5. Keep useful evidence for later evaluation. A completed conversation does not automatically approve its contents for training.

For an external coding agent, follow agent setup and complete the harmless context verification before asking it to execute a workload.

Add a model request to your application

1. Create an API key with only the required scopes.
2. Follow the Inference quickstart, list available models, and choose a catalog ID.
3. Make one bounded request on your server. Keep the key out of the browser and retain the request identity when debugging.
4. Add streaming or tool calling only after checking support for the selected model.
5. Inspect usage and errors before routing more traffic. Direct model calls do not create a Work sandbox; use the SDK when tools and files are part of the task.

Evaluate before improving

1. Create a small representative Dataset, retaining provenance and expected outcomes.
2. Configure a Grader and check it on known passing, failing, and ambiguous examples.
3. Freeze dataset and grader releases, choose a target, and run a bounded Experiment.
4. Inspect per-case results and retained evidence. Separate runtime failures, missing scores, and genuine behavioral failures.
5. Improve the smallest responsible layer, then compare the candidate under the same conditions. Use distinct held-out tasks to assess generalization.

Train and review a Ponder candidate

1. Establish the evaluation baseline above before preparing Ponder training.
2. Review task eligibility, reward binding, runtime readiness, starting model, and spending limit. Keep training and held-out populations separate.
3. Start the reviewed preparation and follow its real run status. Inspect existing state before retrying an interrupted submission.
4. Compare the candidate against the baseline using held-out evaluation. A completed training job is not an acceptance decision.
5. Explicitly accept or reject the candidate, then use Model deployments for a separate deployment action.

Desktop training covers the local workbench path. Continual model training adds recurring approved batches and candidate policy. Recurring Harness review improves the agent's surrounding behavior without changing model weights.