Software agents are increasingly used to answer behavioral questions about code—whether a patch passes tests, which lines caused a failure. These tasks require evidence gathered across the entire repository, but the authors of a new arXiv preprint argue that most existing scaffolds leave intermediate findings in free-form text, which is unstructured and hard to verify.

The paper proposes a "harness-governed" approach to evidence construction, implying a more controlled process for how agents collect and record evidence. The abstract does not provide implementation details, but the significance is clear: structuring evidence could make agentic code analysis more reliable and auditable.

As a preprint, the work has not yet been peer-reviewed, and the abstract is truncated, so only the problem statement and proposed direction are available. Still, the framing highlights a practical gap in current agent scaffolds.