Permitted vs Disallowed Actions in an Agent Policy
Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
Every per-agent rule in an agent-policy.json file divides the world of possible behaviour into two lists: permitted_actions, what the agent may do, and disallowed_actions, what it may not. AGENT-POLICY v0.1 suggests a 12-action vocabulary — from browse and read to checkout, book, and api-call — and one hard law: an action in both lists of the same rule makes the document invalid.
What is an action, in this format?
An action is a verb naming a class of behaviour on your site — not a URL, not a path. That is the format's core break from robots.txt: crawlers are governed by where they may go, agents by what they may do. shop without checkout means "fill a cart, never pay"; read without crawl-full-site means "answer questions from our pages, do not mirror the site". The definitions of each verb are common-sense readings maintained by the spec's vocabulary list, and owners may extend it — unknown action names earn a validator warning, never an error, so the vocabulary can grow without breaking published files.
The twelve suggested actions, grouped by risk
Reading actions — browse, read, search, summarize, quote: the agent consumes content. Acquiring actions — download and crawl-full-site: the agent takes copies away, in bulk in the second case. Transacting actions — shop, checkout, book: the agent moves toward committing its user, with checkout the money step. Writing actions — submit-form and api-call: the agent changes state on your side — posts a form, calls an endpoint. Grouping them this way turns list-writing into risk-writing: most sites permit the first group broadly, price or pause the second, and fence the third and fourth behind named rules and explicit decisions, as the spec samples each demonstrate.
Why two lists instead of one?
Because silence is ambiguous and the format refuses to trade in it. A permitted-only list leaves an agent guessing whether an unlisted action is forbidden or merely unconsidered. A disallowed list resolves the actions the owner cares most about — usually the irreversible ones — into explicit refusals, while the permitted list resolves the welcome ones into explicit grants. What remains unlisted falls back to the rule's verdict and the owner's contact address. The store sample exploits both lists at once: checkout, submit-form, and api-call are disallowed in all three of its rules, so no reading of the file, however creative, finds a path to paying.
The overlap ban is the lists' integrity rule. An action in both lists gives the agent two contradictory instructions from the same rule, and conformance rule 3 declares the whole document invalid rather than letting readers improvise a winner. The free validator checks every rule for exactly this, and the mistakes guide shows how overlaps usually sneak in — typically when one list is edited months after the other.
Writing lists for a real site
Start from journeys, not from the vocabulary. Walk the three or four things an agent might plausibly do on your site — answer a question about a product, compare your prices, book an appointment, file a support request — and map each journey to its verbs. Permit the journeys you welcome, disallow the steps you refuse, and notice how often the refusal is a single verb at the end of an otherwise welcome journey: checkout after shop, submit-form after read. Vertical guides on this site walk the mapping for online stores, booking sites, and SaaS products, where api-call does the heavy lifting.
Worked example: a furniture maker's two lists
Fictional example. Hartwell & Pine, a fictional made-to-order furniture maker, sells 140 pieces with lead times from 3 to 11 weeks. Its policy permits browse, read, search, and quote for every named agent — questions and quotations are the showroom's job. It disallows checkout and book outright: no agent may commit a customer to a 9-week build. shop sits in neither list for its best shopping-agent rule, deliberately — the owner wants cart-building negotiated per agent under conditional contact, not granted by default. In its first quarter, 23 agent-assisted enquiries arrived through the contact route the policy advertises; 0 agent-placed orders existed to unwind, because the two verbs that could have created one were refused in writing.
Splitting the action lists in five steps
- List your agent journeys. Write down what agents plausibly come to do before opening the vocabulary.
- Map journeys to verbs. Assign each journey its actions from the suggested twelve.
- Permit the welcome. Put granted actions in
permitted_actions, per rule. - Disallow the irreversible. Name the refused steps —
checkout,submit-form,api-call— explicitly. - Validate the split. Run the file through the validator; any overlap between the lists is an error to fix, not a nuance to ship.
Keep reading
Rate Limits for AI Agents in agent-policy.json, Explained · Commercial Terms for AI Agents: free, paid-tier, or contact · What Is an Agent Access Policy (v0.1)? · Spec samples · Conformance · Pricing · All guides
Frequently asked questions
If an action is in neither list, can the agent do it?
The lists do not grant by silence. The agent's standing comes from the rule's allowed verdict and the stated terms; an unlisted action is unconsidered territory, and a careful agent treats the owner's contact — or a conditional verdict — as the route to an answer. Owners who care about an action should list it, one way or the other.
Why is overlap an error instead of disallowed winning?
Because the spec refuses to define tie-breakers. First-match-wins and disallowed-wins are both absent from v0.1 by design: a contradiction means the owner has not decided, and the honest machine response is to reject the document so a human decides. The validator enforces this as conformance rule 3.
Can I invent my own action names?
Yes. The vocabulary is non-exhaustive and validators warn — never fail — on unknown action names. Prefer the suggested twelve where they fit, since agents already know them; invent a name only for behaviour none of the twelve describes, and define it in your data_use_terms or notice if it is subtle.
Is crawl-full-site just a heavy read?
No — it is listed separately precisely because it is different in kind: systematic copying of the whole site rather than reading pages to answer a question. The blog sample allows read and quote broadly while its one denied rule disallows crawl-full-site explicitly; many owners permit reading and refuse bulk copying in the same rule.
Which actions do the samples disallow most?
Checkout, submit-form, and api-call — the state-changing and money-moving verbs. The store sample disallows all three in every rule while welcoming browse, read, search, and shop. That pattern, welcome the looking and refuse the committing, is the most common shape in real policies.
Generate a policy — free · Validate a policy · What an agent access policy is