TeleTune learns agent skills from unlabeled user logs without replay

TeleTune's core novelty is handling three specific failure modes when extracting skills from raw telemetry: sparse annotations, no replay access, and procedural knowledge buried in unstructured interaction sequences. The framework doesn't require success/failure labels or the ability to re-run trajectories, which is what makes it applicable to production logs where those signals are often missing or expensive to obtain.

This directly extends the conversation started in 'Mining Agent Skills from Production Traces' (October 5), which tested six combinations of training evidence and found that unlabeled mixed traces can preserve downstream performance when ground truth is unavailable. TeleTune appears to operationalize that finding by providing a concrete method to extract reusable skills from exactly that scenario. The related work on the Agent Error Dataset (September 30) also shares the same premise: production systems generate rich behavioral data that current training pipelines discard. TeleTune and AutoSynthData (October 2) represent complementary approaches to the same economic problem: reducing reliance on hand-labeled or synthetic data by mining what's already there.

If TeleTune's skill libraries transfer to downstream tasks with >80% of the performance of skills trained on goal-annotated data, that validates the core claim. Watch whether the arXiv paper includes cross-domain transfer experiments (skills learned from one software domain applied to a different one) within the next two weeks, as that would signal whether the extracted procedures are genuinely reusable or domain-specific artifacts.