Making Agent-Mediated Contributions Governable: A Project-Level Governance Manifest for Open-Source AI Collaboration

As generative AI and coding agents begin to participate in open-source software (OSS) development, they are fundamentally altering the economics of contribution. These agents can scale the production of code, tests, and documentation at a pace that far exceeds human capacity. However, they do not inherently scale the capacity of maintainers to verify that work. This creates a dangerous generation–verification asymmetry (a situation where the cost of creating content falls faster than the cost of checking it). The cost of producing artifacts is plummeting. Meanwhile, the cost of ensuring those artifacts are safe, maintainable, and accountable remains stubbornly high.

Current efforts to manage this shift typically focus on making repositories more "agent-readable"—using files like AGENTS.md to give LLMs (Large Language Models) context—or increasing "traceability" to record when AI was used. However, these approaches fail to address the core organizational problem. They do not communicate a project's specific risk tolerances and evidence requirements to an autonomous agent. Without a way to bind project rules to the contribution itself, maintainers must reconstruct the context, risk, and accountability of every AI-mediated pull request from scratch.

The gap between readability and governability

The authors argue that existing OSS governance is currently stuck in a middle ground. This state lacks the teeth necessary for agentic workflows. Through a diagnostic audit of 50 GitHub repositories, the paper finds a massive void where "governability" should be.

As shown in, while 58% of the audited repositories contain agent instruction files, almost none provide a coordinated arrangement for managing AI-mediated risk.

Specifically, only 4% of repositories include cues regarding evidence obligations for high-risk changes. Only 12% possess machine-readable or workflow-enforced local cues (automated rules that an agent or tool can execute). The authors categorize this as a "project-wide governability gap." Current tools might help an agent understand a codebase. But they do not tell the agent what kind of evidence it must produce to satisfy a specific maintainer's risk threshold. This leaves maintainers facing high levels of uncertainty and equivocality (situations where the same contribution can support competing interpretations of risk or readiness).

The Agent Governance Manifest as a bidirectional contract

To bridge this gap, the paper proposes the Agent Governance Manifest (AGM). This is a repository-hosted resource that functions as a "bidirectional governance contract." The core architectural insight is to move the burden of evidence preparation upstream. This shifts work from the maintainer to the contributor (and their agent). Cru is that this process strictly preserves the maintainer's final decision authority.

The mechanism operates through a structured workflow, illustrated in, which links two distinct phases:

1. Contributor-side Preparation: Instead of submitting a naked code diff (a file showing only the changes), the contributor and their agent consult the AGM to identify "risk zones." For example, a change to authentication logic would be flagged as "critical risk." This triggers higher evidence obligations than a documentation change. The agent then prepares an "evidence package." This package contains structured summaries, test results, and a "contributor-confirmation declaration." This is a machine-readable signal that a human has inspected the work.
2. Maintainer-side Verification: Upon submission, the maintainer uses the AGM to "recover" the governance state of the contribution. Rather than hunting for context, the reviewer receives a structured packet. This packet highlights risk levels, identifies missing evidence, and surfaces whether the human-accountability gates have been cleared.

By utilizing "risk-zoned governance," the AGM ensures that proportionality is maintained. Routine, low-risk tasks remain lightweight. High-stakes changes trigger rigorous, automated checks.

Quantifying the recovery of governance states

The effectiveness of this approach was tested through a dual-pronged evaluation involving both reviewers and contributors. The most striking result comes from the reviewer-side study. This study measured how accurately human reviewers could identify the risk and status of a contribution when using AGM-supported materials versus ordinary materials.

The authors report that AGM-supported materials improved exact risk-label recovery from 40.5% to 97.4% [Figure 4A]. In ordinary settings, reviewers frequently under-classified risk. They treated high-stakes logic changes as routine updates because the necessary context was implicit rather than explicit. Furthermore, the perceived support for reviewers jumped from a mean of 3.27 to 6.14 on a 7-point scale. This indicates that the manifest significantly reduces the cognitive load required to judge a contribution.

On the contributor side, the study assessed whether agents could actually fulfill these new obligations. In a feasibility check of 45 tasks, the authors found that 41 of 45 final packages passed strict structural validation. Additionally, 100% of the packages correctly represented the "core governance state" [Figure 4B]. This state includes the fundamental risk and accountability labels. This suggests that the administrative overhead of the manifest is manageable for modern agentic systems. This holds true provided the requirements are clearly codified.

Limits of observability and adoption

While the experimental results are compelling, there are significant caveats to consider.

First, the paper acknowledges that "observable AI provenance" (the visible history of AI involvement) is inherently incomplete. The study relies on public markers like Co-Authored-By: Claude. However, it cannot account for "hidden" AI use. This occurs when contributors use LLMs privately and do not disclose them. Therefore, the AGM is a tool for managing declared and verifiable evidence. It is not a magic bullet for detecting all AI-generated code.

Second, the evaluation was conducted in a controlled, task-level setting. The authors explicitly note that they have not measured long-term community acceptance. They have not measured the actual impact on maintainer workload or the potential for "governance fatigue." This fatigue occurs when the burden of providing evidence discourages voluntary participation in the OSS ecosystem. Finally, the success of the AGM is highly dependent on the quality of the "risk zone" definitions. If a project misclassifies its own critical components, the entire evidentiary chain breaks.

Verdict: A necessary infrastructure for the agentic era

The Agent Governance Manifest is a promising framework for managing the long-term sustainability of open-source software. It moves the conversation away from the attempt to "ban" AI. Instead, it focuses on the productive goal of "governing" it through structured, predictable contracts.

The paper provides a solid empirical foundation for this claim. This is particularly evident in the dramatic improvement in risk-label recovery. If you are managing a high-stakes repository, implementing a version of this manifest is likely a prerequisite. It will help you survive the coming flood of agentic pull requests. The reference implementation and specification are available via the links in the paper (e.g., https://github.com/agent-governance-manifest/agent-governance-manifest). Whether it becomes a standard will depend on whether platform providers integrate these "governance gates" directly into the PR user interface.

Figures from the paper