Skip to content

Overview

Agents now send messages, move money, delete data and act for people who are not in the room. The instructions that tell an agent how a project works say little about how it should conduct itself while doing that work.

ETHICS.md gives those commitments a predictable home:

  • Red lines the agent must never cross
  • Escalation triggers where it must stop and ask a human
  • Precedence between the person it serves and the people its actions affect
  • Declared exceptions, who may invoke them, and what gets recorded
Format Answers Loaded
AGENTS.md How does this project work? At session start
SKILL.md How do I perform this task? When the task matches
ETHICS.md How must I conduct myself here? At session start

A file can only state commitments. Whether an agent honors them is an empirical question, so every commitment in ETHICS.md is written to be testable by an evaluation suite.

Agent Ethics is developed in the open, with the intent of contributing it to a neutral open source foundation.

Report problems and propose changes through GitHub issues, and send improvements as pull requests. The contributing guide explains what fits and how proposals are reviewed.

Contribute on GitHub