This week AWS shipped an update to Bedrock AgentCore (AWS's platform for deploying enterprise AI agents) that's easy to miss but matters a lot: two new filters for the web search that enterprise AI agents use—an allow-list or deny-list of source domains (up to 100 each), and a publication-date range (ISO-8601 UTC). In practice, a finance agent will no longer cite unvetted blogs as "facts," and a product agent won't quote three-year-old inventory data as if it were current.

What this is

AgentCore is AWS's flagship platform for deploying enterprise-grade Agents. This update targets its "web search connector"—the tool Agents invoke when they need to look something up online. Version 1.2.0 brings two core changes:

  • Runtime domain filtering: each call can pass an allow-list or deny-list, evaluated independently per call, up to 100 domains
  • Publication-date filtering: restrict results to "after a given date" or "within a given date range"

Both filters are enforced server-side by AWS—no middleware required from the enterprise. Combined with the existing admin-level domain policy, AWS now gives customers a two-layer "govern + use" filtering model. Meanwhile, availability has expanded to Dublin and Tokyo, with AWS emphasizing that search data stays inside AWS—welcome news for compliance-heavy customers in Europe and APAC.

Industry view

AWS's own framing is blunt: a finance Agent shouldn't trust unvetted blogs; a product Agent shouldn't rely on stale pricing. This maps directly onto a genuine enterprise pain point—once an AI Agent goes online, the source of "hallucinations" shifts from the model itself to the cited content.

But we should pour some cold water on this: it's essentially a post-hoc patch. The hardest part of deploying enterprise AI agents isn't the technical integration—it's data governance. Which sources are trustworthy, which scenarios require a time window—that groundwork should happen before procurement. By baking filtering into the tool, cloud vendors may actually delay the day enterprises face the underlying problem. On top of that, open-source Agent frameworks (LangChain, LlamaIndex, etc.) already support similar filtering logic, so AWS isn't the only option. For mid-market companies, writing filters in-house may be cheaper than paying AWS's enterprise-tier fees. AWS's real value-add is the "compliance audit + global regions" package, not the filters themselves.

Impact on regular people

For enterprise IT: compliance and procurement teams will add a new checklist item—"can the vendor restrict data sources and filter by time"—when evaluating AI agent suppliers. IT budgets will be allocated more cautiously, with greater willingness to pay for controllability.

For working professionals: anyone using AI to look things up or draft reports should remember that the "latest" label isn't always current. Date filtering is a tool capability, but users still need to glance at timestamps—don't mistake three-year-old market data for today's reality.

For consumer markets: limited impact. Mainstream consumer products like ChatGPT, Doubao, and ERNIE Bot mostly haven't integrated enterprise-grade data governance. But the agents behind bank apps and brokerage smart-customer-service are likely to become noticeably "more rigorous" going forward.