Define a narrow task
Describe the input, the expected output, and the cases that matter. Request classification or structured extraction is a better starting point than a general promise to “understand the business.”
Choose the data and processing route
Use examples you have permission to use. Identify the approved training backend, model license, hardware, budget, and destination. Ask an Agent with the training capability to organize the workflow.
Prepare and review the examples
Agents can help assemble examples and check their quality. Resolve or set aside disagreement. Hold back a test set that will not be used to train the candidate.
Train and compare
Run the supported training job and compare the candidate with the baseline on held-out work. Review the types of mistakes as well as the overall result.
Make the accepted model available
Deploy to the approved hosting route and select the agents or tasks that may use it. Consulting a specialist and changing an Agent’s main model are separate decisions.
Follow the results
Keep the accepted version connected to its examples and evaluation. Monitor the intended task and use rollback when a previous version is the better choice.
Training is in private preview with eligible Apple silicon, Linux, and separately configured service backends.