Define a narrow task
Describe the input, the output you expect, and the cases that matter. Classifying requests or pulling structured fields out of them is a better place to start than a broad goal like “understand the business.”
Choose the data and processing route
Only use examples you have permission to use. Settle on the approved training backend, the model license, hardware, budget, and destination. Then ask an Agent with the training capability to organize the workflow.
Prepare and review the examples
Agents can help gather examples and check their quality. Resolve any disagreements, or set those examples aside. Hold back a test set that won’t be used to train the candidate.
Train and compare
Run a supported training job and compare the candidate with the baseline on the held-out work. Look at the kinds of mistakes it makes as well as the overall result.
Make the accepted model available
Deploy it to the approved hosting route and choose which Agents or tasks can use it. Letting an Agent consult the specialist is a different decision from changing that Agent’s main model.
Keep an eye on results
Keep the accepted version linked to the examples and evaluation behind it. Watch how it does on the task it was built for, and roll back if an earlier version turns out to be the better choice.
Training is in private preview, on eligible Apple silicon and Linux backends and on separately configured service backends.