AWS Adds Runtime Domain and Publish-Date Filters to AgentCore Web Search

What happened

Amazon Web Services has added runtime domain and publish-date filtering to Web Search on Amazon Bedrock AgentCore, its platform for building and connecting AI agents. The capability ships as part of web-search connector version 1.2.0 and lets developers control, on a per-API-call basis, which web domains an agent may search and what publication-date range results must fall within — all enforced server-side with no external orchestration.

The release also expands Web Search availability to two new AWS regions: eu-west-1 (Dublin) and ap-northeast-1 (Tokyo), giving customers in Europe and Asia Pacific regional endpoints instead of routing traffic across the Atlantic.

The technical details

Connector v1.2.0 introduces two new fields inside the filters object of the Web Search tool input schema:

• filters.domainFilter.include — an allowlist; only results from these domains are returned
• filters.domainFilter.exclude — a denylist; results from these domains are suppressed
• filters.publishedDateFilter.from / filters.publishedDateFilter.to — ISO-8601 UTC bounds restricting results to a specific publication window

Each domain list supports up to 100 entries, counted independently across four possible lists: admin include, admin exclude, runtime include, and runtime exclude. Both filter types are optional; omitting them preserves prior behavior where all indexed content is eligible.

AWS says the filtering layers admin-level policy (set when the connector target is created) with per-call runtime parameters. Include lists merge by intersection — a runtime-requested domain not already on the admin allowlist is silently dropped. Exclude lists merge by union — a domain blocked at either level stays blocked. According to AWS, runtime filters can only narrow scope set by an administrator, never expand it, meaning a caller cannot search a domain the admin hasn't approved or unblock one the admin has denied.

When filters are active, AWS states the system favors precision over recall: results lacking a recognizable domain or publication date are excluded rather than returned unfiltered, which can reduce total result counts.

Web Search is delivered through AgentCore Gateway, AWS's managed, MCP-compatible endpoint for connecting agents to tools. Enabling the feature requires pinning a Web Search target to connector version 1.2.0 or later, IAM permissions (bedrock-agentcore:InvokeGateway and bedrock-agentcore:InvokeWebSearch), and an updated AWS SDK.

AWS also reiterates that AgentCore uses a "zero-egress architecture" where search queries stay within AWS infrastructure — a point the company positions as relevant for regulated customers in the newly added Dublin and Tokyo regions.

What this means

This is a governance feature, not a new model or a new product. It targets a specific pain point for enterprises deploying web-grounded agents: uncontrolled sourcing. A financial-services agent citing an unvetted blog, or a support agent pulling stale documentation, are real liability risks once agents move from demos into production. Server-side enforcement with admin/runtime layering lets platform teams set hard boundaries (e.g., .gov domains only) while still letting individual requests narrow further — useful for multi-tenant systems serving different customers with different compliance needs from one gateway.

The regional expansion to Dublin and Tokyo matters less for capability and more for compliance — EU data-residency requirements and latency-sensitive APAC deployments have been a recurring blocker for enterprises adopting cloud-based agent tooling. Combined, these changes signal AWS is optimizing AgentCore for regulated, enterprise-grade agent deployments rather than adding new intelligence — a incremental but practical step for teams already running production agents on Bedrock.