Study finds coding agents build working software while missing license constraints
A new arXiv paper reports that coding agents often produce software that builds and passes tests yet fails to respect licensing rules, exposing a gap between functional and licensing correctness.
In plain words
Coding agents are AI tools that can write and modify code across a whole software project. The paper's authors tested whether these agents notice and follow software-license rules, which govern how code may be used or redistributed. They found agents often complete the technical work but overlook the legal constraints attached to the code they use.
What was announced
The authors, writing in a paper on arXiv, note that existing evaluations of coding agents prioritize functional success over licensing correctness and commercial suitability. Their study asks whether agents recognize and preserve license constraints when developing proprietary software, covering permissive and copyleft licenses, missing licensing information, transitive dependencies, source modification, dual licensing, vendoring, and distribution contexts.
Experiments across multiple agents evaluated agent traces, repository changes, and final artifacts against licensing ground truth. The results showed a persistent gap between functional and licensing correctness: agents could build and test software while overlooking constraints involving provenance, authorization, and distribution. Outcomes improved when licensing information was translated into concrete engineering actions such as dependency substitution, architectural changes, and release gating. The authors argue that licensing in agentic development requires recognizing and preserving constraints across the software lifecycle, not merely classifying licenses.
What it does not show
- The evidence is limited to the paper's abstract, so the specific agents tested, benchmark sizes, and detailed results are not available in the supplied passages.
- The abstract does not name the agents or quantify the gap between functional and licensing correctness, so no specific figures can be reported.
For security and IT teams
- Scope
- The paper covers repository-level development scenarios including dependency selection, source modification, and software deployment, evaluated against licensing ground truth.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.