01
Acquisition
What result does policy produce for the declared agent identity, resource and proposed method?
Product
A retrieval policy decision point—not another scraper rule parser.
AgentGuard combines supplied source evidence, entitlement records and your own rules at the place where an agent can still change course.
Decision boundary: AgentGuard reports an operational result from your configured policy and the evidence supplied to it. It does not grant a license, determine legal rights, or replace legal review.
The problem
Your policy is currently spread across code, tickets and memory.
A crawler rule answers a narrow acquisition question. A public usage signal expresses a source preference. A contract record may evidence a scoped entitlement. Your organization—not AgentGuard—determines the legal basis for use.
AgentGuard keeps those layers distinct, preserves where each supplied claim came from and reports the strictest applicable result produced by your configured policy.
Decision model
Five dimensions. One executable result.
There is no universal “safe” flag. Each decision shows how the independent dimensions resolved and why.
02
Processing & use
Do attached signals address the declared purpose, retention, transformation and output?
03
Entitlement evidence
Does a caller-attested record claim to cover this resource, use and time?
04
Tenant policy
What does your organization require when evidence is absent, stale, conflicting or restrictive?
05
Review state
Should ambiguity stop automation and route to a named policy owner?
Modular target architecture
Decision logic in the cloud. Enforcement where the response exists.
The pilot provides scoped API access, server-stored policy revisions, tenant decision history, usage limits and exports. Your runtime supplies normalized evidence. Managed fetching, live enforcement and signed decision receipts require separate integrations.
Customer runtimeBrowser · crawler · RAGDeclares identity and intended use
request→
Policy enforcement point · plannedSDK · sidecar · proxyHolds the response until a final decision
decision + evidence↔
AgentGuard data planePolicy decision APIDeterministic evaluation over frozen inputs
Evidence adapters · plannedEntitlement registry · plannedDecision record signing · planned
Named policies, guided templates, synthetic tests and owner-approved immutable revisions. Explicit key assignments activate a revision without removing workspace safeguards. Usage and filterable decision history remain tenant-scoped.
Authenticated checks and deterministic evaluation over caller-supplied normalized evidence.
A separately scoped customer integration would capture the response delivered to its identity, region and network.
Source-handling model
Evaluate what is requested—not an imagined map of the whole internet.
The planned collector minimizes cost and liability by collecting relevant published signals at point of need, then caching only where repeated customer traffic justifies it.
ModeHandlingStatus
Caller suppliedEvaluate bounded normalized evidence without fetching the URLCurrent pilot
On demandCollect relevant public policy signals for new or expired resourcesPlanned collector
Selective cacheReuse domain/path evidence only within an explicit freshness windowPlanned production beta
Source packPre-warm customer-nominated high-value domains—not the whole webDemand-led enterprise option
Adopt without breaking traffic
Observe first. Enforce deliberately.
The current pilot starts with sanitized replay traffic. Live shadow and enforcement are customer-integration options to validate separately before any result can control a pipeline.
Measure the policy before it controls the pipeline.
- Compare decisions to current behavior
- Find missing evidence and contract scopes
- Backtest a new policy revision
Convert reviewed decisions into runtime actions.
- Block or hold on explicit reason codes
- Apply retention and attribution constraints
- Route to customer-configured licensing review
Evidence, not magic
Every claim keeps its source and attachment point.
Evidence familyWhat it contributesBoundary
RFC 9309 robotsCrawler/path acquisition rulesNot access authorization
Usage signalsPurpose-specific source preferencesVocabulary support varies
Response metadataResource-specific headers and head tagsVisible only after retrieval
RSL / licensing signalMachine-readable use terms or a licensing requirementNo verified offer URL or entitlement
Entitlement recordCustomer-attested scope and obligationsCaller-supplied in pilot; verification is planned
Tenant policyYour organization’s enforcement postureCannot invent publisher permission
See the semantics
Change the purpose. Watch the decision change.
The sandbox runs eight frozen scenarios and makes no request to their display URLs.