Free plan

← All guides

Rate Limits for AI Agents in agent-policy.json, Explained

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

The short answer

The rate_limit field in an agent policy is an optional, per-rule politeness cap written as {"requests_per_minute": N}. It asks an agent to make no more than N requests per minute on your site. The published samples use 60 for a favoured shopping agent, 30 for a reading agent, and 10 for the catch-all — a ladder of trust expressed in one number per rule. null, or omitting the field, states no limit.

Generate your policy — free · Validate an existing one

What is a policy rate limit?

It is a requested cap, declared by the site owner inside the rule that grants an agent access — the speed limit sign on a road the agent is invited to drive. It is not a server throttle: nothing in the file slows anyone down mechanically, and the spec is explicit that a policy declares rather than enforces. Its power is contractual, not technical. An agent that accepts your terms accepts your number; an agent observed ignoring it has broken a stated term, which is a very different conversation from "you were going too fast" with no number ever published. Server-side rate limiting remains your infrastructure's job — see policy vs bot blocking for where the line sits.

Reading the samples' ladder: 60, 30, 10

The store sample's three numbers encode three relationships. 60 requests per minute for openai/agent: a named shopping agent on a paid tier, expected to work a catalogue briskly. 30 for perplexity/agent: a reading and quoting agent on free terms — half the speed, because reading to answer questions needs fewer fetches than working a store. 10 for generic/*: strangers who have not yet made contact get a walking pace. The pattern to copy is not the numbers but the reasoning: rate follows the job the rule permits and the trust the rule expresses. A rule permitting crawl-full-site — which the format lets you disallow instead, and most owners do — would need a different conversation entirely about bulk access and commercial terms.

Picking your own numbers

Start from your ordinary traffic, not from fear. If your site comfortably serves a few hundred human page views a minute, a 60-per-minute agent cap is a rounding error; on a small site where 60 requests a minute is the traffic, it is not. Consider the permitted actions: search-heavy rules burn requests faster than read-only ones, and an agent comparing 200 products will legitimately fetch more than one answering a single question. Consider, too, that the cap is per rule — three agents at 60 per minute is 180. If you cannot pick, copy the samples' ladder as a starting position, publish it, and revise with evidence: the field is one number in one rule, and last_updated exists to date the revision.

Two failure modes to avoid. Setting the cap so low that permitted work becomes impossible — an agent allowed to shop at 1 request per minute cannot compare a category — reads as a refusal wearing a welcome's clothes. And stating a cap you never look at again: pair the published number with whatever server logs you already keep, because the comparison between declared terms and observed behaviour is where the field earns its keep, as the ignores-your-policy guide explains.

Rate limits and the rest of the rule

The cap does not travel alone. It sits beside commercial_terms — the samples pair the highest cap with the paid tier, which is a coherent offer: pay, and work faster — and beside data_use_terms, which govern what the agent may do with what it fetched at that speed. A complete rule reads as one bargain: verdict, actions, terms, pace, data. The free generator asks for the rate limit per agent rule in the same pass as the rest, so the bargain is written as a whole rather than patched later.

Worked example: a ticket marketplace tunes its ladder

Fictional example. FrontRow Exchange, a fictional resale ticket marketplace with 18,000 live listings, launched with the samples' ladder copied verbatim: 60, 30, 10 across its 3 rules. Listing pages turn over fast, and within a month its logs showed the 30-per-minute reading agent exhausting its cap by 09:20 on on-sale days while the 60-per-minute shopping agent rarely passed 22. The owner swapped the emphasis, not the philosophy: the reading agent rose to 45 with search permitted, the shopping agent settled at 40, and the catch-all stayed at 10. Nothing else in the file changed; last_updated moved, and the two affected operators were emailed the new numbers the same day — the kind of boring, dated maintenance that keeps declared terms credible.

Setting rate limits in five steps

  1. Price the job, not the fear. Estimate the requests the permitted actions genuinely require.
  2. Ladder by trust. Named partners fastest, known readers next, the catch-all slowest.
  3. Sum the rules. Add the caps together and check the total against what your site comfortably serves.
  4. Pair pace with terms. Let commercial terms and rate limits tell one story per rule.
  5. Publish, observe, revise. Compare declared caps with observed behaviour and update the number — and the date — when they diverge.

Keep reading

Commercial Terms for AI Agents: free, paid-tier, or contact · data_use_terms Explained: Retention, Resale, and Training · How to Tell AI Agents What They May Do · Spec samples · Conformance · Pricing · All guides

Frequently asked questions

Is rate_limit required in a rule?

No. It is one of two optional rule fields, alongside data_use_terms. Omit it, or set it to null, and the rule states no limit. Omitting is a decision too: agents reading the rule learn that the owner set terms for actions and data but expressed no preference on pace.

Does the rate limit block agents that exceed it?

No. The policy declares the cap; it cannot enforce it. Mechanical throttling lives in your server, CDN, or firewall. What the published number gives you is a stated term: an agent over the cap is not merely fast, it is outside the terms it was offered, which matters for every conversation that follows.

Per minute per agent, or per minute total?

Per rule — that is, per agent identifier. Each named agent gets the cap written in its own rule. If four rules each allow 30 requests per minute, the file as a whole contemplates up to 120 requests per minute of agent traffic, plus whatever the default brings.

What number should a small site start with?

The samples' ladder is a sane default: 60 for a favoured named agent, 30 for a reader, 10 for the catch-all. A very small site can halve all three. The right test is whether an agent can complete the work the same rule permits — a cap that makes permitted actions impossible contradicts the rule that contains it.

Can different actions have different limits in one rule?

Not in v0.1. The rule carries a single requests_per_minute figure covering the agent's activity under that rule. Finer-grained pacing — by action or by path — is outside the format today, one of the deliberate omissions the spec lists for future versions.

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

← Back