Choose a specialist.
Open Agent Academy and select your agent. Its allowed tools remain permission-controlled.
StarloOS 0.9.0 adds a separate Train an agent screen. It learns a bounded document classifier from labelled examples, evaluates separate held-out examples, and abstains on weak matches. This is distinct from the prompt-teaching Academy described below; neither trains a foundation model.
Local workbench guide →Agent Academy is implemented in the local desktop edition. It stores approved context and runs practice tasks through your connected model.

Open Agent Academy and select your agent. Its allowed tools remain permission-controlled.
Save a policy and examples of the answers you expect. New missions receive a snapshot of that teaching revision.
Save a practice task with required and forbidden phrases. Run it using your own model connection and inspect the actual result.
Approved instructions and examples are supplied to future agent tasks. Memory and attached knowledge provide additional context. You decide what to save; there is no automatic, invisible learning from all conversations.
Phrase checks can show whether specific words appear. They cannot prove factual correctness, reliability, or that an agent will behave correctly in every situation. Practice can consume paid provider usage. This feature is not model fine-tuning and does not update model weights.
This page is the public walkthrough. Interactive Academy editing is available inside the local desktop runtime, not as a public unauthenticated training service on this website.