Batch Validating Agent Policies: 100 Files per Call
Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
Batch validation runs the AGENT-POLICY v0.1 conformance checks across many policy documents at once. AgentGate's batch API accepts up to 100 policies per call (a Pro feature) and returns the same verdict structure as the single validator — errors and warnings, per file. It exists for the moment one policy becomes many: a fleet owner's quarterly sweep, an agency's portfolio check, a platform verifying the policies its customers publish.
What is batch validation?
It is the single validator's discipline at fleet scale. The checks are identical — the five conformance rules, required fields and domains, overlaps, duplicates, formats — because conformance does not change with quantity. What changes is the workflow around them: instead of a human pasting one document into a form, a process fetches or gathers the current documents and submits the set, receiving a per-file report it can triage. One honesty note travels with the scale: like the single validator, the batch API validates documents you bring to it — making no network requests of its own, per the spec's design — so fetching the live copies remains your pipeline's first step, and the right one.
Who needs it, and when
Three operators, three rhythms. Fleet owners — several sites, one organisation — run it at every review sweep and after every release train, because releases are how live policies quietly break. Agencies run it monthly across client portfolios, as in the agency pattern, where the batch report becomes a standing section of client reporting. Platforms and hosts run it as a service-health instrument: customers' published policies sampled or swept, breakages surfaced before counterparties find them. The common trigger is identical in all three: the number of files exceeded what pasting into a form can honestly cover — usually somewhere around five — and the cost of a missed invalid file started exceeding the cost of the pipeline.
The sweep workflow, step by step
Fetch. Pull each origin's live /.well-known/agent-policy.json — the live copy is the only one that matters, as the validation guide insists — and keep the fetched set as the sweep's evidence. Submit. Send the set to the batch API, up to 100 documents per call; larger fleets chunk by hundreds, which is an afternoon's scripting, not an architecture. Triage. Work the report in conformance order: errors first (a file the world is reading is invalid now), then staleness (compare each file's last_updated against your register), then warnings to be decided. Fix at the source. Corrections happen in each site's own publishing flow — the batch report is a diagnosis, not an editor — followed by a confirmation run over just the repaired files. Record. The sweep's counts and exceptions go into the register, dated; that record is what turns validation from an event into a control.
What batch scale changes — and what it must not
Scale tempts two dilutions, both worth refusing. The first is sampling: checking a subset because the whole set feels heavy. At 100 files per call, a 400-policy fleet is 4 calls; sampling is a false economy that converts a control into a survey. The second is verdict drift: treating fleet-level averages ("96% valid") as health, when conformance is per-file and binary — the 4% are specific sites whose counterparties are reading invalid documents today, and the report names them. Batch validation's value is precisely that it keeps per-file honesty affordable at scale: every file, every sweep, with the same five rules a single shopkeeper's policy faces.
Worked example: a hosting platform sweeps 214 customer policies
Fictional example. Northbeam Hosting — the same fictional host from the signature audit — sweeps the 214 customer sites on its platform that publish agent policies, in 3 batch calls, on the first Monday of each month. A typical month's report: 206 valid, 5 with errors (most often an overlap introduced by a customer's edit, once a platform update that re-serialised files without the signature field), 3 unreachable at fetch time (2 migrations, 1 expired domain the host could flag to its own customer). Error owners receive the specific findings the same day; the platform's fix-confirmation run closes the loop within the week. The sweep costs the platform an hour of automation time and produces the sentence its enterprise customers keep quoting back: your policy estate, checked monthly, by the numbers.
The batch validation checklist
- Fetch live copies. The sweep validates what the world reads — well-known URLs, fetched at sweep time, kept as evidence.
- Submit the whole set. Up to 100 per call; chunk larger fleets rather than sampling them.
- Triage in conformance order. Errors, then staleness against the register, then warnings for decision.
- Fix at the source, re-run the repairs. Batch reports diagnose; each site's own flow corrects; a confirmation run proves it.
- Record the sweep. Counts, exceptions, and repairs into the register — the control is the record, not the run.
Keep reading
Hosted Agent Policy URLs: When AgentGate Serves Your File · What Is the agent-policy.json File? · Agent Access Policy Examples · Spec samples · Conformance · Pricing · All guides
Frequently asked questions
How many policies can one call check?
Up to 100 policies per call on AgentGate Pro. Fleets larger than that chunk their submissions — a 350-file estate is four calls — which keeps the workflow a simple loop rather than a platform project. The single-policy validator remains free and unlimited for individual checks.
Does the batch API fetch policies from URLs itself?
No. Following the spec's design, the validator makes no network requests: you bring the documents. In practice the sweep pipeline fetches each origin's well-known URL first — which is a feature, since the fetched copies are the sweep's evidence and the fetch itself catches unreachable policies before validation begins.
Are batch results the same as single validation?
Yes — the same conformance rules, the same error and warning classes, reported per file. Scale changes the delivery, not the standard. A file that passes in a batch would pass alone, which is what makes the sweep's per-file verdicts actionable rather than statistical.
Who is batch validation for?
Anyone whose policy count outgrew the paste-into-a-form workflow: fleet owners with several origins, agencies holding client portfolios, platforms and hosts watching customer estates. The practical threshold is low — once a sweep means opening more than about five files by hand, the batch call is the honest way to keep the whole set checked.
What should happen when a sweep finds errors?
Treat them as live defects with owners: the named site's owner gets the specific findings the same day, fixes through their normal publishing flow, and the repair is confirmed by a follow-up run over just the affected files. The register records the exception and its closure. An error that survives two sweeps is no longer a bug — it is a governance finding.
Generate a policy — free · Validate a policy · What an agent access policy is