Supervised fine-tuning on agent trajectories is a known way to improve coding agents, but this arXiv preprint (2610.23134) argues that picking useful training examples is not just a matter of keeping successful runs. A resolved patch, the authors note, may still rest on inefficient reasoning that makes it a poor teaching signal.
To address this, the paper introduces MetaCurate, which evolves curation filters for high-quality agentic coding data. Rather than applying a fixed rule for what counts as a good example, the filters themselves are refined — though the abstract is truncated, so the mechanics of that evolution are not described in the available text.
Because there is only one source, there are no conflicting findings to reconcile. The main claim stands on its own: success alone is a weak proxy for data quality, and curation needs to look beyond whether a patch resolved the task.