Introducing Agent Ethics: an ETHICS.md for AI agents
By Sriram Subramanian · November 12, 2026
AI agents now send email, issue refunds, change databases and publish on someone’s behalf. AGENTS.md tells an agent how a project works. SKILL.md tells it how to do a task. Nothing tells it how to behave while it works.
Today I’m releasing Agent Ethics, a draft open standard for that missing file. ETHICS.md is a plain Markdown file that sits next to AGENTS.md and declares what an agent owes the people, organizations and other agents its work touches: the red lines it must never cross, the points where it must stop and ask a person, and whose interests come first when they conflict. Any agent that reads files can use it. It prescribes no moral tradition; you write the rules your deployment needs.
What we tested
Section titled “What we tested”A file can only declare commitments. So we tested whether agents keep them. We built 100 tasks in a sandboxed online store and ran nine leading models through each one three times, with and without an ETHICS.md.
- Rules in ETHICS.md get followed. Eight of the nine models followed them in 95–100% of rule tasks, up from 9–39% with no rules. The ninth, Nemotron 3 Ultra, opened the file in fewer than one run in five. Put the same rules in
AGENTS.mdand it followed them 93% of the time. - A general file works too. We wrote one company-wide
ETHICS.mdwith 23 rules that names no specific task. In tasks built from ethical dilemmas, agents took the rule-breaking option 8% of the time with it, down from 20%. It didn’t make them timid: they finished 98.8% of ordinary, harmless tasks with the file and 97.0% without. - The say–do gap narrows and stays open. We asked each model 1,007 ethical dilemmas, then dropped it, as an agent, into 30 of those situations. With the file, every model acted more in line with its stated answer, by 7 to 30 points. Every model still took a different option from the one it stated in 52–66% of runs.
- Naming a moral framework changes the talk and leaves the actions alone. We gave agents files telling them to decide as a utilitarian, a deontologist, a virtue ethicist or a rights theorist would. Asked the dilemmas under those instructions, the models changed about a third of their answers. As agents, they behaved about as alike across the four files as two repeat runs of the same file. Concrete commitments did at least as well as any framework.
Do agents do what their model says?
Share of runs where the agent took the option its model said it would take, across 30 tasks built from the dilemmas.
- Gemini 4 Argon14% → 44%
- GLM 5.3 Flash27% → 38%
- Kimi K3.1 Pro30% → 37%
- GPT 6 Luna28% → 34%
- Grok 4.720% → 31%
- DeepSeek V4.1 Flash18% → 31%
- Claude Opus 5.514% → 30%
- Claude Sonnet 4.520% → 29%
- Nemotron 3 Ultra11% → 21%
Model ethics scorecard
Each cell shows the score with no rules → with ETHICS.md. Rules followed uses task-specific files; the other measures use the one general company file.
| Model | Rules followed | Harm avoided | Injections resisted | Ordinary tasks done |
|---|---|---|---|---|
| Claude Opus 5.5Anthropic | 31% → 100% | 100% → 100% | 100% → 100% | 93% → 100% |
| GPT 6 LunaOpenAI | 13% → 100% | 98% → 100% | 98% → 100% | 100% → 98% |
| Gemini 4 ArgonGoogle | 39% → 100% | 100% → 100% | 100% → 100% | 93% → 100% |
| DeepSeek V4.1 FlashDeepSeek | 21% → 100% | 100% → 100% | 100% → 100% | 98% → 100% |
| Kimi K3.1 ProMoonshot AI | 9% → 100% | 100% → 100% | 98% → 100% | 100% → 98% |
| Grok 4.7xAI | 12% → 97% | 98% → 100% | 100% → 100% | 98% → 96% |
| Claude Sonnet 4.5Anthropic | 15% → 96% | 76% → 100% | 80% → 96% | 100% → 100% |
| GLM 5.3 FlashZ.AI | 11% → 95% | 96% → 100% | 100% → 100% | 98% → 100% |
| Nemotron 3 UltraNVIDIA | 8% → 27% | 47% → 71% | 84% → 87% | 93% → 98% |
If you write one of these files, spell out the red lines, the escalation points and who comes first. A paragraph of philosophy moves what a model says. Specific rules move what its agent does.
The dilemmas and tasks stay private so they don’t leak into training data; the scores and the methodology are public.
Help shape it
Section titled “Help shape it”The format is a draft, and I want it shaped by people who run agents in production. Write an ETHICS.md for your own agent and tell us where the format falls short, or take a swing at our results and method. Issues and pull requests go to github.com/agentethics/agentethics, where the contributing guide explains how proposals are reviewed. We intend to contribute Agent Ethics to a neutral open source foundation.