Free plan

← All guides

Agent Policy vs Terms of Service: Which Does What?

Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team

The short answer

Your terms of service and your agent policy are not rivals; they address different readers in different languages. Terms of service are legal prose for humans — customers, counterparties, courts. An agent policy is a structured JSON declaration for software: which agents may operate, what actions, on what terms, fetchable at /.well-known/agent-policy.json. A site with agent traffic needs both, each doing the job it is shaped for, and neither pretending to be the other.

Generate your policy — free · Validate an existing one

What is the difference, in one line each?

Terms of service: the contract-shaped document governing the human relationship with your service — accounts, liability, acceptable use, disputes. Agent policy: the machine-readable declaration of your terms for agent operation specifically — verdicts, action lists, commercial terms, pace, data use — in the AGENT-POLICY v0.1 format. The overlap is real (both state terms) and the distinction is functional: one is read by people and lawyers, at human speed, in dispute-shaped contexts; the other is read by agents, at fetch time, before the behaviour it governs happens. The comparison mirrors agent policy vs robots.txt — three documents, three audiences, one site.

Why the terms of service cannot do the policy's job

An agent planning a task cannot act on a prose document. It needs to know, mechanically: am I permitted, which verbs, at what pace, under what data conditions? A terms-of-service clause saying "automated access may be restricted" answers none of those — it does not name the agent, the actions, or the terms, and its ambiguity is deliberate legal craft, exactly wrong for an operational reader. Worse, prose terms are usually written without agents in mind at all: silent on identifiers, silent on data after-life, silent on pace. The agent policy exists because the questions agents ask are specific, and specificity is what structured fields are for. Publishing one does not weaken the prose document; it gives the prose document an operational companion that speaks first, at the moment of behaviour.

Why the policy cannot do the terms' job

The spec disclaims it in its first section: a policy is the owner's unilateral declaration, not legal advice, and a validator's valid verdict is not a legal opinion. The file has no machinery for the things terms of service exist to settle — liability, governing law, dispute process, the customer relationship itself. An owner who treats a conforming agent-policy.json as legal coverage has confused a well-built notice with a contract. The sane relationship between the documents is referencing, not replacing: terms of service can point to the policy as the operative statement of agent-access terms; the policy's notice can point back to the terms as the governing document. Each cites the other; neither impersonates the other.

Keeping the two consistent

Consistency is the working risk, because the documents are edited by different people on different clocks. The failure to avoid: terms of service prohibiting all automated access while the policy welcomes named agents — or the reverse, a legal update quietly revoking what the policy publicly grants. The practice that prevents it is procedural: any edit to either document triggers a read of the other. Owners with a policy review cadence — quarterly suits most — add the terms cross-check to the same appointment. When the two genuinely conflict, the resolution is a decision, made by the humans who own both documents, and then written into both — never left for an agent or a court to discover independently.

Worked example: a SaaS company aligns two documents that disagreed

Fictional example. Formstacker, a fictional form-building SaaS with 22,000 customer accounts, published a welcoming agent policy in January — named agents may read documentation and api-call its public status endpoints. Its terms of service, last revised by counsel two years earlier, prohibited "all automated access" without exception. The conflict surfaced in March, when a partner's counsel, doing diligence, quoted both documents in one email. The fix was a coordinated revision: the terms gained a carve-out — automated access as expressly permitted by the published agent policy — and the policy's notice gained a pointer to the terms as the governing document. One afternoon of alignment, after which the company's two voices said the same thing in their two languages, which is all consistency ever is.

The two-documents checklist

  1. Keep both. Terms of service for the human legal relationship; agent policy for machine-readable agent terms.
  2. Cross-reference. Terms point to the policy for agent-access specifics; the policy's notice points back to the terms as governing.
  3. Edit in pairs. Any change to one document triggers a read of the other — same week, same decision-maker.
  4. Never let the policy impersonate legal advice. Valid means conformant; your legal position lives in the terms and your counsel's work.
  5. Audit at review time. The quarterly policy review includes one question for the terms: do we still agree with ourselves?

Keep reading

Agent Policy vs API Keys and Authentication · Agent Policy for Online Stores: a Practical Pattern · What Is an Agent Access Policy (v0.1)? · Spec samples · Conformance · Pricing · All guides

Frequently asked questions

If my terms of service ban bots, do I still need an agent policy?

The ban in prose does not answer an agent's operational questions, and agents are not crawlers — the audiences differ. If your position is refusal, the agent policy states it in the agent-facing form: default block, like the bank sample, cross-referenced from your terms. If your position has any nuance — named partners, reading allowed — only the policy can express it where agents read.

Which document wins if they conflict?

That is a legal question about your specific documents, and this guide — like the spec itself — offers no legal opinion. Operationally, the answer is to never let the question arise: conflicts are drafting failures with a procedural fix. Align the documents, in writing, with counsel where the terms are concerned.

Can the agent policy replace a robots.txt exclusion too?

No — three documents, three audiences. robots.txt governs crawler fetching by path; the agent policy governs agent operation by action and terms; terms of service govern the human legal relationship. Most considered sites publish all three, cross-referenced, each current. The robots.txt comparison guide maps the first pair in detail.

Should my terms of service mention the policy by URL?

Yes — the well-known URL is stable by design, and citing it in the terms does two jobs: it tells human readers (partners, counsel, platforms) where the operative agent terms live, and it binds the documents together as one position stated in two languages. A one-line reference is enough.

Does publishing an agent policy change my legal exposure?

The spec's position is the safe one to repeat: the policy is a unilateral declaration of terms, not legal advice, and conformance is not a legal opinion. Whether and how published terms affect any legal question depends on jurisdiction and facts far beyond a format specification — which is precisely why the terms of service and your counsel remain in the picture.

Generate a policy — free · Validate a policy · What an agent access policy is

← Back