On August 19, 2026, AWS added domain and publication date filters to Web Search on AgentCore. If your AI agents pull data from the web, you can now limit their searches to specific sites and time windows.
The new controls arrive in Web Search connector version 1.2.0, announced on the AWS Machine Learning Blog. It works per request. You can narrow the search scope for a single task without changing settings for everyone else.
Direct control over sources and dates
The update brings two main tools:
- Domain filtering. You can pass
domainFilter.includeanddomainFilter.excludelists. Each list supports up to 100 domains, letting you build custom allowlists and denylists. - Date ranges. The
publishedDateFilterusesfromandtofields in standard ISO-8601 UTC format. This makes time-based queries clean and exact across time zones.
AWS handles this filtering on the server side. The search service drops blocked sites before sending results back to your application. The release also adds support for two new regions: eu-west-1 in Dublin and ap-northeast-1 in Tokyo.
Where this helps in daily work
These controls make targeted searches much easier. For compliance, you can restrict agents to approved industry portals. For market intelligence, you can compare sources between two exact timestamps. If you serve multiple clients, you can apply tenant-specific domain rules with ease. Customer support bots can also focus on recent product documentation.
What AWS claims versus reality
AWS presents zero-egress, regional data proximity, and lower latency as key benefits of this architecture. These numbers and speed claims come from AWS marketing — nobody outside the company has tested them independently under real workloads. A domain filter also does not stop prompt injection inside approved pages. It controls scope, not safety. Scope is not truth.
How to test it in your stack
Before moving to production, run a controlled test in your region. Pin connector version 1.2.0 and test the same query under four conditions: no filter, an allowlist, a denylist, and a date window. Keep your test variables constant.
Check every returned URL, publication date, blocked domain, citation, and latency number. Watch for cross-region data transfers, and test how your model handles messy text from allowed sources. Once your tests confirm the filters work as expected, you can safely deploy.