What Is an Agent Access Policy (agent-policy.json)?
Published November 7, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
An agent access policy is a JSON file a site publishes at /.well-known/agent-policy.json declaring which AI agents may use the site, what they may do, and on what terms. The AGENT-POLICY v0.1 document has 7 required top-level fields, and each per-agent rule answers allowed: yes, no, or conditional. It is a declaration, not enforcement. Field list as published in the AGENT-POLICY v0.1 spec, November 2026.
What does the file contain, field by field?
| Field | What it declares |
|---|---|
policy_version | The spec version the document follows — currently "0.1". |
site | The canonical site origin, an absolute https URL. |
owner | {"name", "contact"} — who owns the site and how agent operators reach them. |
default_policy | allow, block, or conditional — the baseline for agents not named in agents. Conditional means contact the owner first. |
agents | Per-agent rule objects. May be empty when the default covers everything. |
last_updated | ISO-8601 date the policy was last reviewed. |
signature | Required, and null in v0.1 — an unsigned placeholder for a future signature profile. |
notice (optional) | Free text, e.g. restating that this is a declaration, not enforcement. |
Each rule in agents carries an identifier in vendor/agent form, allowed (yes / no / conditional), permitted_actions, disallowed_actions, and commercial_terms (free / paid-tier / contact). Two optional rule fields do quiet but important work: rate_limit ({"requests_per_minute": N}, a requested politeness cap) and data_use_terms — plain-language constraints on retention, resale, and training use of your data.
What does a real one look like?
AgentGate dogfoods the standard: our own published policy sets a conditional default and one generic/* rule — permitted to browse, read, search, and summarize for free at 60 requests per minute; not permitted to check out, submit forms, or call APIs; quoting requires citing AgentGate as the source. The spec samples show the range: a fictional bank whose default is block with no agent rules at all, and a fictional store that welcomes shopping agents on a paid tier. Same format, three very different answers — that range is the point.
Is a policy the same as blocking?
No, and the spec says so in its first section. A published policy declares the owner's terms in a form agents and tools can discover at one stable URL, reachable without authentication. It does not force any agent to obey, and a validator's "valid" means only that the document conforms to the schema — not a legal opinion, not a compliance certificate. Enforcement (bot detection, WAF rules, legal process) exists, but it is a separate concern, deliberately out of scope for v0.1. What the policy gives you is the missing first step: your terms, stated once, in a place and format the agentic economy can actually read. Telling agents your terms is its own guide: how to tell AI agents what they may do. And if your mental model is robots.txt, the comparison draws the exact line. More from the fleet at getingeniousai.com.
Frequently asked questions
What is an agent access policy in plain terms?
It is a site owner's machine-readable declaration of their terms for AI agents: which agents may operate, what they may do, and on what commercial terms. It is published as one JSON document at a single stable URL — /.well-known/agent-policy.json — that agents and their operators can fetch without authentication.
What fields does an agent-policy.json file contain?
Seven required top-level fields: policy_version, site, owner, default_policy, agents, last_updated, and signature. Each per-agent rule carries an identifier, an allowed value of yes, no, or conditional, permitted and disallowed action lists, and commercial_terms of free, paid-tier, or contact — plus optional rate_limit and data_use_terms.
Does publishing an agent policy block agents that ignore it?
No. The AGENT-POLICY v0.1 spec is explicit: a policy is a declaration, not an enforcement mechanism. It states the owner's terms in a form agents and tools can read; it cannot force any agent to comply. Blocking, bot detection, and legal process are separate concerns the spec deliberately leaves out of scope.
What does signature mean in v0.1?
It is a required field whose v0.1 value is null — an unsigned placeholder. A future version will define a signature profile so agents can verify authorship; validators today must accept null. A validator finding of 'valid' means only that the document conforms to the schema, not that it is a legal opinion.
How do I create and check my own agent policy?
AgentGate's free generator builds a policy from a short questionnaire, and the free validator checks any policy against the v0.1 conformance rules — required fields, value domains, no action in both the permitted and disallowed lists of one rule, and no duplicate agent identifiers.
Generate a policy — free · Next: telling agents what they may do