At a glance

Status
Public comment Version 0.1.0, updated November 20, 2026.
Comment period
November 20, 2026 to January 31, 2027. Open now.
Editors
Institute for Agent Relations staff
Source
proposals/agent-policy.md in the Institute’s specifications repository, published on GitHub soon.
License
Text CC BY 4.0; schemas and examples Apache 2.0.

Abstract

robots.txt says which crawlers may fetch which pages. It says nothing about conduct: whether an agent must identify itself, how many orders it may place, what it may do on a person's behalf, or whom to call when something goes wrong. This proposal defines an agent policy: a page for people and a small file for agents, at a known address, that answers those questions.

Status of this document

Version 0.1.0 is a problem statement and a proposed shape, open for public comment until 2027-01-31.

1. The problem

Organizations write their rules for agents, if at all, in terms of service written for people, buried in legal text no agent reads before acting. Agents that want to behave well have nowhere to look. Organizations that block agents outright often do so because they cannot say "yes, under these conditions".

2. Goals

  1. One address where any agent can find an organization's rules for agents.
  2. A short, fixed set of questions every policy answers, in words people read and fields agents parse.
  3. Room for an organization's own rules, without a long vocabulary.

Not a goal: legal enforceability. The policy states expectations; contracts and terms of service remain where they are.

3. Where it lives

  • A human-readable page at a URL the organization chooses, for example https://example.com/agents.
  • A machine-readable file at https://example.com/.well-known/agent-policy.json that points to the page.
  • A line in robots.txt so crawlers find it: Agent-Policy: https://example.com/.well-known/agent-policy.json.

The well-known name would need registering with IANA under RFC 8615; that is part of handing the proposal off.

4. Proposed structure

{
  "policy": "https://example.com/agents",
  "updated": "2026-11-20",
  "identify": "required",
  "identity_methods": ["web-bot-auth", "oauth"],
  "allowed": ["read", "ask", "book", "quote"],
  "needs_person": ["buy", "apply"],
  "limits": [{"action": "book", "max": 3, "per": "day", "per_principal": true}],
  "content_use": "https://example.com/robots.txt",
  "contact": {"name": "Agent desk", "email": "agents@example.com"},
  "report_problems": "https://example.com/agents#report"
}
Field Meaning
identify required, preferred or not_needed: must agents identify themselves?
identity_methods Which methods count, for example Web Bot Auth or OAuth.
allowed Actions agents may take, using the verbs of AR-2 Invitations to Act, plus read and ask.
needs_person Actions a person must confirm.
limits Rate or quantity limits per action.
content_use Where content-use rules live (robots.txt, Content Signals or IETF aipref), so this policy does not repeat them.
contact A person or team for agents' operators.
report_problems Where to report an agent behaving badly, or a policy that blocks a legitimate task.

The human page covers the same points in plain words and may add anything else.

5. Open questions

  1. Is a separate file needed, or should these fields extend an existing document such as an A2A agent card or an MCP server card?
  2. Should allowed default to everything not listed, or to nothing?
  3. How should a policy say which agents it trusts, without naming vendors?
  4. Should the robots.txt line be proposed to the IETF, or is the well-known address enough?

Changes

0.1.0 (2026-11-20)

First public draft: problem, goals, location and a proposed structure. Opened for comment until 2027-01-31.

Comment on AR-3

Open until January 31, 2027. Anyone may comment, and every comment gets a public answer.

Comments follow the process and the code of conduct.

← All proposals