- The paper identifies repository context files such as AGENTS.md as a new governance layer where developers translate values including fairness, accessibility, sustainability, inclusivity, and accountability into instructions for AI coding agents.
- The paper’s preliminary review of 25 AGENTS.md files highlights six examples and distinguishes vague ethical aspirations from executable constraints, such as named-data fairness tests that can guide measurable agent behavior.
- The paper proposes research on large-scale mapping, developer translation practices, compliance experiments, longitudinal change, and interactions between natural-language directives and programmatic agent configurations.
Overview and motivation
"Operationalizing Ethics for AI Agents: How Developers Encode Values into Repository Context Files" (2605.05584) is a vision paper by Treude, Baltes, and Cheong that examines an emerging practice in agentic software engineering: developers writing ethical directives into repository-level context files such as AGENTS.md. The paper's central argument is that a new object of study has appeared — machine-interpretable, natural-language constraints authored by humans and interpreted by AI coding agents during development workflows — and that this layer constitutes a developer-authored governance mechanism distinct from prior Responsible AI frameworks aimed at human developers.
The authors position their contribution against two research traditions. The first is the operationalization of ethics and values in software engineering: professional codes of conduct, Value-Sensitive Design, Value-Based Engineering, IEEE 7000, SLEEC requirements, Ethical User Stories, and the Responsible AI Pattern Catalogue. They note recurring critiques that such codes "lack mechanisms to reinforce their own normative claims" and provide limited actionable guidance for everyday design decisions, with systematic reviews showing weak support for later lifecycle stages. The second tradition concerns AI agents themselves, spanning agent-based social simulation ethics, taxonomies of human–AI collaboration, the CRAFT values framework for agentic SE, trustworthiness dimensions for AI software engineers, and recent empirical studies of context files — including analyses of 2,303 "Agent READMEs" across 1,925 repositories, 466 open-source projects characterizing context expression styles, and a survey of eight configuration mechanisms across 2,853 repositories identifying AGENTS.md as an emerging interoperable standard used in over 60,000 open-source projects.
Preliminary investigation
The empirical component is deliberately preliminary: the authors searched GitHub and used ChatGPT web search to locate candidate repositories, manually inspected 25 AGENTS.md files, and selected six illustrative examples. They are explicit that this sample motivates a research agenda rather than providing a comprehensive account. Even so, the examples substantiate the paper's core claim that developers translate abstract principles into situated directives:
| Repository |
Encoded value |
Directive form |
| ez-recipes-web |
Accountability, transparency |
Human-in-the-loop, accountability for changes |
| awesome-copilot |
Fairness |
Structured bias tests with named test data |
| DHCP |
Tone |
Prohibition on moral lectures or unsolicited opinions |
| haxtheweb/create |
Accessibility, sustainability |
Accessibility maximization; environmental and community sustainability |
| home-assistant/core |
Inclusivity |
Non-discriminatory language; clarity for non-native English speakers |
| magentaA11y |
Respectful language, bias awareness |
Verification-oriented responses |
A notable contrast the authors draw is between vague and executable encodings. One repository instructs the agent to "follow all guidelines for ethical AI" without specifying triggers, conditions, or enforcement mechanisms — documenting intent but not behavior. Another embeds fairness as evaluation logic, supplying explicit test names (e.g., 'John Smith', 'José GarcÃa', 'Lakshmi Patel') and directing the agent to detect differential outcomes for equivalent qualifications. This distinction between declarative aspiration and executable constraint is arguably the paper's most analytically useful observation, since it anticipates the compliance question that dominates its roadmap.
Research agenda
The roadmap identifies five directions. First, large-scale empirical mapping of context files to determine which ethical categories dominate (bias, accessibility, sustainability, privacy) and how emphases vary across domains and cultures. Second, studying the translation process itself: how developers decide what to encode, at what granularity, and how these constraints are negotiated in pull requests and between human and AI contributors, using repository mining combined with qualitative analysis. Third, experimental evaluation of whether agents actually adhere to encoded constraints — comparing agent behavior with and without context files on bias mitigation, tone, accessibility, and regulatory compliance — which would distinguish genuine behavioral shaping from mere signaling of intent. Fourth, longitudinal analysis of how encoded ethics evolve over time, including strengthening after incidents, relaxation under productivity pressure, and the possibility that some directives amount to box-ticking rather than substantive commitment. Fifth, extending analysis beyond natural-language files to programmatic configuration artifacts — agent skills, tool permissions, hook scripts, system prompts — and asking whether these reproduce, complement, or contradict prose directives.
The authors also raise a conceptual question: whether ethics communicated to machines should resemble human-oriented documentation such as codes of conduct, which rely on shared norms and contextual judgment, or whether fundamentally different machine-oriented representations are required.
Limitations and open questions
The paper is candid about its evidentiary limits. The six-repository sample was selected for illustrative value, so no claims about prevalence or distribution of ethical content across context files can be made; indeed, related work cited in the paper suggests security and performance considerations are rarely specified in Agent READMEs, implying ethics-related content may be uncommon. Several substantive questions remain open: whether high-level directives like "follow all guidelines for ethical AI" produce any measurable behavioral change in agents; whether agents can resolve the tension between "doing the right thing" and "doing things right" when guided only by natural-language repository text; and whether conflicting directives across contributors' edits are negotiated or silently resolved. The authors also acknowledge that ethical principles may be too vague, unenforceable, or mutually contradictory — a problem inherited directly from human-oriented ethics frameworks.
Conclusion
This vision paper reframes AI governance as an empirically tractable software engineering problem: rather than debating agent ethics in the abstract, researchers can study the concrete artifacts through which developers already encode fairness, accessibility, sustainability, tone, and privacy into agent-facing instructions. Its contribution lies less in findings than in identifying a new artifact class and articulating a coherent agenda — mapping, translation dynamics, compliance testing, longitudinal evolution, and full-configuration-space analysis — for grounding AI governance in observed development practice.
