Microsoft's TeleTune Evolves Agent Skills From Offline Usage Logs

Microsoft researchers have published TeleTune, a method that improves an agent's software skills using only raw offline usage logs, with no live test environment.

According to the paper as reported by Rohan Paul, TeleTune guesses each logged session's goal and has the model predict every logged action using a text skill library. When a goal guess is wrong, it suggests edits to the library, and an edit is kept only if next-action prediction accuracy rises on held-out logs.

The reported claim is that this held-out next-action accuracy tracks live success, which is what removes the need for a live environment. Earlier methods such as Agent Workflow Memory, by contrast, need goal-labeled examples or a live environment, according to the same report. These claims come from the paper's summary as relayed in the report, not from independent testing.