What If an AI Agent Ignores Your Policy?
Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
Then the agent ignores it — and nothing in the file stops it, because AGENT-POLICY v0.1 is explicit from its first section: a policy is declaration plus discovery, not enforcement. Publishing terms does not block a single request. What the declaration still does, even against an ignoring agent, is real: it puts your terms on the public record in standard form, which changes every conversation, complaint, and enforcement action that follows. Enforcement itself lives elsewhere — in your server, your firewall, and the law.
What does "declaration, not enforcement" mean?
It means the format claims one power and renounces another. The claimed power: stating your terms, in a standard machine-readable document, at the address agents check, in values agents can act on. The renounced power: compulsion. No field in agent-policy.json reaches into an agent's runtime. The spec refuses the enforcement claim on purpose — a format that pretended to block would be lying, and its "valid" verdicts would inflate into promises they cannot keep. The same honesty governs AgentGate's own surfaces: a validator's valid means schema conformance only, never that agents obey or that the law is satisfied.
What the declaration does against an ignorer
Three things, none of them nothing. It fixes the record. Before the policy, your terms were an inference from your site's design; after it, they are a dated document an operator can be shown. "Your agent checked out against a published disallow" is a sentence with evidence attached. It sorts the ecosystem. Agents and operators who respect declared terms — the ones building reputations on it — identify themselves by complying; ignoring a clear, conforming, easily fetched policy is a choice that says something about its maker. It lowers every other cost. Hosts, platforms, payment providers, and counsel all move faster when the terms were published in standard form than when they must be reconstructed from screenshots and recollection.
Where enforcement actually lives
In the layers the spec names as out of scope — which is not the same as unimportant. Server-side rate limiting and bot management act on behaviour regardless of declarations. Web application firewall rules can fence the paths behind your disallowed verbs. Authentication walls protect what was never public. Legal process addresses the operator behind the agent. Policy vs bot blocking draws the working line between the declaration layer and these enforcement layers, and the sane architecture uses both: the policy tells good-faith agents the rules, and the enforcement layers handle the rest — each doing the job it is actually capable of.
Your own behaviour matters to the declaration's weight. A policy that is valid, current, and consistently stated is a record worth pointing to. A stale, contradictory file undermines the complaint it would support — one more reason the maintenance habits in the review-cadence guide are part of enforcement in everything but name.
What not to do about an ignorer
Do not respond by making the policy dishonest — inventing enforcement claims in the notice, or treating the file as a security control in your own risk assessments. Do not abandon the declaration because it is not a wall; locks and notices do different work, and houses need both. And do not escalate inside the format: there is no sterner field to set, no "block harder" value in v0.1, and pretending otherwise wastes the effort that belongs in the layers that can actually act. Set the named agent's rule to no if it has one — the refusal belongs on the record — and then enforce in the places enforcement exists.
Worked example: a wholesaler's ignored policy still pays off
Fictional example. Pacific Rim Components, a fictional electronics wholesaler, published a policy refusing crawl-full-site for every agent and capping its catch-all at 10 requests per minute. One agent mirrored its 26,000-line catalogue anyway, at production-disturbing pace. The policy blocked none of it — and was still the most useful document in the response. The wholesaler's host acted within a day on the strength of the published terms it was shown; the operator, contacted through the channel the policy itself advertises, apologised and pointed to a misconfigured client that had never fetched the file. The catalogue crawl stopped; the operator's other agents began fetching the policy first. The firewall rule the wholesaler added that week handles the next ignorer mechanically — and the policy, unchanged, remains the reason anyone can say what the rule enforces.
When an agent ignores your policy
- Verify the record. Confirm your live file is valid and current — the declaration must be clean before it can be pointed to.
- Document the behaviour. Logs showing the actions your policy disallows, at the pace it refuses, are the evidence the declaration frames.
- Contact the operator. Use the same channel your policy advertises; many ignorances are misconfigurations, and the record makes the conversation short.
- Enforce in the right layer. Rate limits, firewall rules, and host or platform abuse processes do the mechanical work.
- Record the refusal. Set the agent's rule to
nowhere it has one — the public record should say what happened.
Keep reading
Should You Let AI Agents Shop on Your Site? · Let Agents Browse but Not Check Out: the Split Rule · Agent Access Policy Examples · Spec samples · Conformance · Pricing · All guides
Frequently asked questions
If a policy cannot stop anyone, why publish one?
Because stopping is not the only work terms do. A published policy tells good-faith agents — a large and growing share — exactly how to behave on your site, converts your terms into a dated public record, and gives hosts, platforms, and counsel something standard to act on when an agent misbehaves. Walls without notices punish everyone; notices without walls rely on goodwill. Serious sites run both.
Does ignoring a policy break the law?
That is a legal question the format deliberately does not answer. The spec states a policy is the owner's unilateral declaration, not legal advice, and a validator's verdict is not a legal opinion. What ignoring a clear published policy does, independent of law, is create a documented fact pattern that lawyers and platforms can evaluate.
Can I make my policy enforceable by adding fields?
No. v0.1 defines no enforcement fields, and inventing some in your file gives you a document no agent is built to honour mechanically. Enforcement is a property of your infrastructure — authentication, throttling, firewall rules — not of the declaration. Keep the file honest and put the machinery where machinery works.
What should I do first when I spot an ignoring agent?
Check your own file. Fetch the live well-known URL and validate it: a surprising share of 'ignored' policies turn out to be stale, invalid, or unreachable copies that the agent never actually received. If the file is clean and current, document the behaviour and contact the operator with the record in hand.
Will a future version add enforcement?
The spec's stated roadmap concerns authentication of the publisher — a signature profile — not enforcement of terms. The declaration/enforcement split is architectural, not a missing feature: documents declare, infrastructure enforces, and v0.1's successors are expected to keep that line where it is.
Generate a policy — free · Validate a policy · What an agent access policy is