Free plan

← All guides

How Often Should You Update Your Agent Policy?

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

The short answer

Update your agent policy on two clocks. The event clock: edit the file the day a decision changes — an agent graduates, a term is renegotiated, a fence moves. The review clock: re-read the whole file on a schedule — quarterly suits most sites — and move last_updated even when the review changes nothing, because the date certifies the review happened. A policy outside both clocks becomes the zombie document every guide on this site warns about.

Generate your policy — free · Validate an existing one

What does "updating" a policy mean?

Three distinct acts share the word. Editing changes content: rules, verdicts, numbers, terms. Reviewing re-decides the existing content against current reality and may change nothing. Republishing puts the result at the well-known URL and validates the live copy. The format's last_updated field records the first two — the spec describes it as the date the policy was last reviewed — and the versioning guide draws the consequence: the date moves on reviews as well as edits, because freshness is a claim about attention, not about diffs.

The event clock: changes that cannot wait for the quarter

Four events should edit a policy the day they land. A relationship resolves: a conditional negotiation concludes — write the named rule, in whichever direction it resolved, while the agreement is warm. A relationship sours: an agent observed breaking stated terms gets its refusal recorded; the file is the public record, and the record should be contemporaneous. The site changes: a new storefront, a new booking engine, a paywall — any change that moves your action surface moves the policy that describes it. The ecosystem changes: a significant new agent appears and deserves a named rule rather than life under the catch-all. Each event edit follows the same liturgy: change, date, validate, republish, validate the live copy.

The review clock: what a quarterly review actually does

Ninety minutes, four questions. Accuracy: does every rule still describe a relationship that exists, at terms both sides would recognise? Retire the dead, repair the drifted. Coverage: who has been living under the catch-all — and does the quarter's contact traffic say someone graduates, or someone gets named in refusal? Coherence: do the fences still match the site — the verbs your journeys actually use, the after-life your licences require? Freshness signals: is the contact staffed, the notice still true, the version posture still right? Then the date moves, whether or not a single rule did. Owners of several policies run the review as a fleet sweep, per the multi-site guide; the Team tier's change history exists to make what changed since last review a lookup rather than an archaeology.

Too often is also a failure mode

Policies are read by counterparties who build on them. A file whose terms churn weekly teaches operators that your declarations are weather, not climate — and agents caching copies will perpetually hold versions you have already replaced. The discipline that prevents it: batch deliberate changes into the review where they can wait, reserve event edits for the four triggers above, and treat term changes affecting named partners as communications, not just commits — the operator hears it from you before their agent fetches it from the file. Stability is a feature of a declaration; the review clock exists so that stability never has to mean staleness.

If your organisation keeps exactly one artefact from this guide, let it be the review appointment itself — recurring, owned, and short. Policies rarely die from a dramatic wrong decision; they die from eighteen months of small unrecorded drift, discovered by a counterparty. The appointment is the whole defence, and its modesty is the point: four questions, a date moved, a log entry written. Everything else — graduation speed, negotiation quality, fence precision — improves as a side effect of the file being looked at by its owner on a rhythm, because drift cannot accumulate where attention is periodic. Set the first appointment before you publish the first version, and let the calendar, not your memory, be the policy's maintenance department.

Worked example: a publisher's year of policy maintenance, in numbers

Fictional example. The Marigold Review, a fictional literary publisher with 640 essays and reviews online, kept a simple log of its policy's first year. Event edits: 6 — three agents graduated from conditional to named rules, one agent's rule was set to no after repeated rate-limit breaches, one new booking-adjacent verb was fenced after a site redesign, one partner's pace doubled by agreement. Scheduled reviews: 4, of which 2 changed nothing but the date. Total edits: 11, average 20 minutes each including validation. Total surprises at review time: 0 — the log's real product. The publisher's editor summarises the system in one sentence: the policy never drifted more than a fortnight from the truth, because the truth kept appointments with it.

The update-rhythm checklist

  1. Edit on events. Resolved negotiations, broken terms, site changes, significant new agents — same-day edits, dated and validated.
  2. Review on a schedule. Quarterly for most sites: accuracy, coverage, coherence, freshness — then move the date regardless.
  3. Keep a change log. Your own register, or Team-tier change history, so each review starts from what changed, not from memory.
  4. Batch the deliberate. Non-urgent term changes wait for the review; counterparties build on your file's stability.
  5. Communicate partner changes. Named agents' operators hear term changes from you before their agents fetch them.

Keep reading

Agent Policy vs Terms of Service: Which Does What? · Agent Policy vs API Keys and Authentication · Agent Access Policy Examples · Spec samples · Conformance · Pricing · All guides

Frequently asked questions

What is the minimum viable review rhythm?

Quarterly suits most sites: frequent enough that drift stays small, sparse enough that the review is a real event rather than a chore to fake. High-change environments — marketplaces, fast-moving catalogues, many named partners — justify monthly. Annual is neglect wearing a schedule's clothes: too much ecosystem moves in a year.

If nothing changed, should last_updated still move?

Yes. The field records when the policy was last reviewed, and a review that re-decides the existing terms is a real event — it is the difference between 'these terms were true once' and 'these terms were stood behind this quarter'. Agents and operators read the date as exactly that signal.

Who should own the review?

Whoever owns the relationships the policy describes — partnerships, digital, or the site owner personally on small sites. The review needs authority to change verdicts and terms, not just to proofread. The ownership guide covers the organisational patterns; the essential property is a named human with the file in their calendar.

Do I need to notify agents when the policy changes?

Their agents will fetch the new file in due course — that is the design. But for named partners whose terms changed materially, direct notice is the professional norm and cheap insurance against a partner operating on a cached copy. The file is the record; the email is the relationship.

Can I automate the review reminders?

The reminder, yes — calendar, ticket, or your host's monitoring. The review itself, no: its content is judgement about relationships and drift, which is the part that must stay human. Automation's proper role here is fetching and validating copies — batch validation for fleets — so the human hour is spent deciding, not checking commas.

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

← Back