Skip to content

The nearest neighbour now shares our venue and our retired word, for a different unit

# research(2026-09). Two papers by the same author attribute agents from pull requests, and both are close enough to RQ1 that a reviewer will raise them.

Fingerprinting AI Coding Agents on GitHub (Ghaleb, MSR '26, arXiv:2601.17406) reads 33,580 PRs from Codex, Copilot, Devin, Cursor and Claude Code, builds 41 features over commit messages, PR structure and code characteristics, and reports 97.2% macro F1 from XGBoost for "identifying the submitting agent". AgenTag (arXiv:2608.00966) extends the same unit to an open-world setting, "a multimodal, learned, open-world problem rather than a closed-set classification task", at weighted F1 0.96 and AUC 0.84 for unseen agents.

What they are not. The unit is the tool that wrote the change, never the organization that reviewed it. The first paper's only use of our unit is as a threat: "We do not test generalizability to private repositories or other platforms (e.g., GitLab), where agent behavior may differ due to stricter style guides or organizational practices." Organizational practice is the confound they cannot control, which is the thing this study measures. Their evidence is hand-built features over commit and PR metadata read by a tree ensemble, not what a model learns from review content, and neither paper reports a true-positive rate at a fixed false-positive rate: macro precision, recall and F1 in the first, F1 and AUC in the second. The first reports Claude Code at 57% recall with 82% precision and does not say what the operating point costs, which is the reading this study refuses to leave out.

What this changes here. Nothing in the design, three things in how it is written.

  1. The related-work section gains these two as the nearest neighbours in our own venue, and the contrast is stated as unit, evidence and operating point rather than as novelty by absence.
  2. Retiring "fingerprint" was decided on other grounds and is now also forced: in MSR '26 the word names agent attribution, and using it for organizational house style would collide with a paper the same readers have just seen.
  3. AgenTag's future work names the closest anyone comes to our question and leaves it open: "quantify how repository-specific instruction files (e.g., CLAUDE.md, AGENTS.md) affect behavioral fingerprints and attribution accuracy." Project-level configuration as a modifier of attribution is one step from organizational convention as the thing attributed, and nobody has taken it.