How to Block All AI Agents (and Say So Clearly)
Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
Blocking every AI agent takes the shortest conforming policy in the format: default_policy set to block, an empty agents list, and the other required fields filled honestly. That is the published bank sample's entire posture — Meridian Trust Bank, fictional, blocks all agents in a handful of lines. This guide covers when the blanket refusal is the right answer, how to write it so it reads as a decision rather than an oversight, and what it does not do.
What does blocking all agents mean in a policy?
It means your published answer, for every agent named or unnamed, is no. The default does all the work because there are no exceptions to state: no rule grants anything to anyone. Note what kind of "no" this is — the standing declaration of the format. The file states refusal in the standard place so that agents built to respect declared terms have a clear instruction to respect, and operators have a clear record. It is not a technical barrier; enforcement machinery is a separate layer, addressed below and in policy vs bot blocking.
When the blanket is the honest answer
Three kinds of sites reach it legitimately. Regulated and sensitive services — the bank sample's logic — where agent operation raises obligations the owner cannot discharge for unnamed software. Sites whose content is the product and whose owners decline agent-mediated consumption in every form, after weighing the demand they are also refusing. Sites that are not ready: no staffed contact, no reviewed terms, no capacity to sort agents — for whom a clear block today beats a neglected conditional that promises conversations nobody will have. The common thread is deliberateness. A block chosen for a stated reason is a policy; a block copied from a sample without the reason is a placeholder wearing a policy's clothes.
Writing the refusal well
The mechanics are minimal, so the craft is in the framing fields. The notice is where a blanket block explains itself: regulated obligations, a licensing-only model with the contact named, a temporary posture with a review date — whatever is true. owner.contact matters even here: operators who believe they have a case for exception need a door to knock on, and several real relationships start as exceptions argued against a block. If exceptions exist or emerge — one agent under contract, one partner — they are written as named yes rules inside the block's file, at which point the policy is a guest list with a closed door, the pattern the default guide describes. Keep last_updated honest: a block reviewed quarterly is a stance; a block untouched for years is an heirloom.
What the block does not do
It does not stop a determined agent at the network level — nothing in the file can, per the spec's honest scope — and it does not speak to crawlers at all, which remain robots.txt's audience; many block-posture sites publish both files saying their respective noes. It also does not end the subject: agents will still arrive, operators will still write, and the ecosystem will keep changing the value of the refusal. The block's job is clarity, and clarity compounds: a site that has clearly said no is in a stronger position in every subsequent conversation — with operators, hosts, and platforms — than a site whose position must be guessed from its silence.
Hold the block to the same standard of honesty as every other posture in these guides: it should be reviewed against reality, not inherited from the week it was written. Postures calcify in both directions — welcomes that outlive their welcome, and blocks that outlive their reasons, still refusing a channel the business would now price and fence like everyone else. The quarterly question for a block is therefore not "is it still safe?" (it is definitionally the safest posture) but "is it still true?" — does it still describe what the organisation would decide today, knowing what the agent economy has become since the file was drafted? A block that survives that question annually is a strategy. A block that never faces it is a habit with a JSON file attached.
Worked example: a credit union's block, with one argued exception
Fictional example. San Pedro Community Credit Union, a fictional member-owned institution with 11,000 members, published the bank pattern almost verbatim: block default, empty agents list, and a notice attributing the posture to member-data obligations, with the digital-services address as contact. In 8 months the contact received 5 exception requests. Four were declined with the same reason the notice states. The fifth — a financial-aggregation agent its members were actively requesting — was evaluated for a quarter and admitted as the file's single rule: yes for read on account-summary pages only, submit-form and api-call disallowed, 15 requests per minute, terms by contract. The block had done its work precisely: it made the exception a decision, with a file that shows both the rule and the one argued breach in it.
The block-all checklist
- Choose it for a reason. Regulation, product, or readiness — be able to state which, in one sentence.
- Write the minimal file correctly. Block default, empty agents list, all seven required fields present and valid.
- Explain in the notice. The reason, the contact, and any review horizon — inside the document that travels.
- Keep a door. A monitored contact turns a wall into a position; exceptions get argued there, not in your server logs.
- Review the refusal. Move
last_updatedwhen you re-decide — a block is a decision with a date, like every other verdict.
Keep reading
How to Welcome AI Agents on Your Site · Agent Policy for Price-Comparison Agents · What Is an Agent Access Policy (v0.1)? · Spec samples · Conformance · Pricing · All guides
Frequently asked questions
Is block with an empty agents list really a complete policy?
Yes — it is one of the three published spec samples. default_policy block answers every agent; with no exceptions to state, the agents array is legitimately empty. The remaining required fields — version, site, owner, last_updated, signature — must still be present and valid.
Why publish a block instead of just staying silent?
Silence is ambiguous; a block is information. Agents built to respect declared terms need the declaration to respect. Operators evaluating your site get an answer without a meeting. And if your terms are ever disputed, a dated public refusal in standard form is a record silence cannot provide.
Does a block policy stop agents technically?
No. Like every policy, it declares rather than enforces. Technical prevention lives in your own stack — bot management, firewall rules, authentication — and a block-posture site should run those layers as its real barrier, with the policy as the public statement of the position they implement.
Can I block all agents except one partner?
Yes — that is a block default plus one named yes rule: a closed door with a guest list. The default refuses the world; the rule admits the partner on stated terms. It is one of the most common real-world postures for B2B sites, and the format expresses it in two lines.
Should the block also go in robots.txt?
The files govern different audiences: robots.txt addresses crawlers and fetching, agent-policy.json addresses agents and actions. A site refusing automated access generally publishes its refusal in both, each in its own vocabulary. Neither file substitutes for the other — see agent policy vs robots.txt.
Generate a policy — free · Validate a policy · What an agent access policy is