Govern the Repository, Not Just the Agent
mlj.solutions
mlj.solutions Studie na 930 000+ agenty vytvořených pull requestech měří integration friction, náklad na zapojení příspěvku do kódu, který mění i ostatní přispěvatelé. Přibližně polovina variace v tomto tření zůstává vlastností repozitáře i po plné kontrole autora, velikosti změny a agenta. U příspěvků od AI agentů se tohle tření navíc koncentruje zhruba dvakrát víc než u lidských (intraclass correlation 0,30 vs. 0,16) ve stejných repozitářích, se stejnými pravidly. Závěr autorů: měřit jen úspěšnost jednotlivého agenta nestačí. Riziko je z velké části na úrovni ekosystému, ne komponenty. 🔗 Govern the Repository, Not the Agent: Measuring Ecosystem-Level Risk in AI-Native Software 🔹 Sledujete nás pro více zajímavostí ze světa AI!
View on Instagrammlj.solutions
mlj.solutions Každý AI pull request projde testy. Projekt se přesto může pomalu rozpadat. Integrační problémy se u AI změn koncentrovaly téměř 2× víc než u lidských. Nestačí měřit dokončené tickety, ale musíte sledovat zdraví celé codebase. 🔵 Jak měříte dlouhodobý dopad AI změn na váš repozitář?
View on InstagramWhat the study is about
Autonomous coding agents now open and merge pull requests in shared repositories at scale, yet they are evaluated one at a time on isolated tasks. A study of more than 930,000 pull requests shows that the risk accumulates in the repository, not in the individual agent.
Statistics
930,000+
pull requests analyzed
These are contributions made by agents in shared repositories, on which the author measured integration friction.
50%
of the variation stays in the repository
Roughly half of the differences in integration friction belong to the repository, even after accounting for the author, the size of the contribution and the specific agent.
2x
stronger tie to the repository
For agent contributions, friction is tied to the repository roughly twice as strongly as for human contributions.
What this means
- Passing the tests isn't enough, agents that pass their own tests still leave repositories with problems that can't be attributed to any single contribution.
- The risk belongs to the ecosystem, the problem lies not in the agent itself but in the repository where contributions pile up.
- Integration friction is measurable, it is the cost of merging a change into code that other contributors are modifying in the meantime.
- Agents concentrate friction more than humans, the difference holds even after accounting for code size, code age, task type and process maturity.
- The whole ecosystem needs governing, AI-native software is better measured and managed at the repository level than one agent at a time.
Our take
Evaluating an AI agent only by whether it passes tests isn't enough. When deploying agents into code, we therefore also monitor what happens to the repository as a whole.
Michal Dobrovolný
Founder & CTO
Zdroj
Russo, D. (2026). Govern the Repository, Not the Agent: Measuring Ecosystem-Level Risk in AI-Native Software. arXiv:2606.28235. https://arxiv.org/abs/2606.28235
Don't miss these posts
When More Agents Pay Off and When They Don't
More AI agents don't necessarily mean a better result. A study shows that what matters is how much information is lost when it is passed between agents.
AI Remembers a Claim Even When You Tell It It's False
When a model is fine-tuned on documents that deny a claim, it often learns the claim as true anyway. The authors named this effect Negation Neglect.
The Skill That Explains Anything to Anyone
ELI5 is an open-source skill for Claude Code that adapts explanations to the audience. With the skill it passed 10 of 12 criteria, without it only 5.
Want to implement AI in your company?
We'll go through your needs together and propose an approach that delivers measurable results.
Write to us
