Making Agent-Mediated Contributions Governable: A Project-Level Governance Manifest for Open-Source AI Collaboration
This paper addresses the governance challenges posed by AI-generated contributions in open-source software by introducing the Agent Governance Manifest (AGM), a repository-hosted framework that significantly improves risk assessment accuracy and review support while preserving maintainer authority through a structured contract between contributor-side evidence preparation and maintainer-side verification.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the internet as a giant, ever-growing city built entirely by volunteers. In this city, the "buildings" are software programs, and the "blueprints" are open-source code that anyone can read, fix, or improve. For decades, this city has run on a simple rule: if you want to change a building, you hand your plans to the city's architects (called "maintainers"), who check them to make sure they are safe and won't cause the whole block to collapse. But now, a new kind of worker has arrived: the AI coding agent. Think of these agents as super-fast, tireless robots that can draft blueprints, write construction manuals, and even fix leaks in seconds. The problem is that these robots can build so fast that the human architects can't keep up with checking their work. It's like having a factory that produces 1,000 house plans a minute, but you only have one inspector who can check one plan an hour. If the robots make a mistake, the whole city could be in trouble. This is the big question this paper tackles: How do we make sure these robot-built plans are safe, accountable, and ready for the human architects to approve, without slowing down the robots or overwhelming the humans?
The researchers behind this study, Jinjin Gao and her team, decided to investigate how open-source projects are currently handling this robot revolution. They looked at 50 popular software projects on GitHub (a giant online library for code) to see if the "city rules" were ready for AI workers. They found that while most projects have good rules for humans and even some instructions for robots on how to read the blueprints, they are completely missing the rules for what the robots must prove before a human can sign off. It's like having a factory with instructions on how to use the machines, but no checklist for safety inspections before the product leaves the line.
To fix this, the team invented a new tool called the Agent Governance Manifest (AGM). Think of AGM as a smart, digital "safety passport" that every robot-generated plan must carry. Instead of just handing over a pile of code, the robot (guided by the human) has to fill out this passport. The passport says: "Here is what we changed, here is the risk level (Low, Medium, or High), here is the proof that we tested it, and here is a human signature saying, 'I checked this, and I take responsibility.'"
The team tested this idea in a controlled experiment. They asked 15 people to act as the city architects and review 75 different "construction plans." Half the plans were just the usual messy code (the "ordinary" way), and the other half came with the new AGM safety passport. The results were clear: when the architects had the AGM passport, they were able to correctly identify the risk level of the changes 97.4% of the time (37 out of 38 cases). Without the passport, they only got it right 40.5% of the time (15 out of 37 cases). The researchers also found that the people felt much more supported by the AGM system, rating it 6.14 out of 7 compared to just 3.27 for the old way.
On the other side of the table, they tested if the robots and their human partners could actually make these passports. They gave 15 different pairs of humans and robots 45 tasks to complete. The result? Every single final package correctly represented the core safety rules, and 41 out of 45 passed strict structural checks. The humans were able to confirm the work, and the robots did the heavy lifting of gathering the evidence, without the robots ever taking over the final decision-making power.
The paper argues that we don't need to ban AI or make the robots do everything. Instead, we need a "governance infrastructure" that sits between the robot's speed and the human's judgment. The AGM acts like a bridge: it forces the robot to organize its work into a clear, evidence-based format that the human can quickly verify. This keeps the human in the driver's seat for the final "yes" or "no," while letting the robot handle the busy work of gathering the proof.
In short, the study suggests that by giving robots a structured "safety passport" to fill out, we can keep the open-source city safe and fast. It's not about stopping the robots; it's about making sure they know exactly what they need to show the human architects before the construction crew starts building. The researchers found that this approach works well in controlled tests, making it much easier for humans to spot risks and verify safety, but they note that real-world testing in live communities is still needed to see how it holds up over time.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.