Source: AWS News Blog — 2026-08-20
Summary
AWS updated Web Search on Amazon Bedrock AgentCore Gateway — its managed, MCP-based grounding tool for agents — to support per-call domain include/exclude lists (up to 100 domains) and published-date range filtering, alongside expansion to two new regions (Europe/Ireland and Asia Pacific/Tokyo). The feature lets a developer constrain what an agent's web-search tool call is allowed to pull in, at the level of an individual request, rather than relying on prompt instructions alone to keep an agent's grounding sources trustworthy and current.
Key Takeaways
- Domain filtering: up to 100 domains can be included or excluded per search call, letting teams restrict an agent to trusted sources (or block known-unreliable ones) at the API level rather than hoping the prompt is followed.
- Date-range filtering: agents can constrain results to a specific published-date window per call — directly useful for any agent workflow (like this digest's own recency rule) that needs to exclude stale sources.
- Delivered as an MCP-based connector on Bedrock AgentCore Gateway — another example of MCP becoming the standard interface layer even for AWS's own first-party tooling, not just third-party integrations.
- Regional expansion to Europe (Ireland) and Asia Pacific (Tokyo) broadens where this managed grounding tool can be used with lower latency and to meet data-residency requirements.
- Signals a broader trend: as agentic RAG matures, "search the web" is evolving from a blunt tool call into a parameterized, governable primitive — closer to a database query than a generic action.
Reel Script
Hook: Most "AI web search" tools will happily ground an answer in a five-year-old blog post or a source you'd never trust — AWS just shipped a fix that's almost embarrassingly simple: let the developer filter by date and domain, at the API level.
Core Concept: When an agent has a "search the web" tool, the quality of its answer depends entirely on what it's allowed to pull back — a RAG system, retrieval-augmented generation, is only as trustworthy as its retrieval step. Historically, controlling that meant writing careful prompt instructions like "only use reputable sources" and hoping the model actually complies, which it often doesn't under pressure to just answer the question. AWS's update moves that control out of the prompt and into the API call itself: domain filtering (an allowlist or blocklist, up to 100 domains) and date-range filtering are now parameters on the search request, enforced by the tool, not suggested to the model.
Hands-On: Picture the difference in practice. Before: your agent's system prompt says "prefer results from the last two weeks and only from official vendor blogs" — and the model sometimes just... doesn't follow that, especially on an edge case. After: the search call itself carries domains: [trusted-list] and a published_after date parameter, so results outside those bounds never even reach the model to reason about. It's the same shift databases went through decades ago — moving validation from "hope the application code checks it" to "the database itself won't return or accept bad data." Delivered as an MCP connector on Bedrock's AgentCore Gateway, this is also a small data point in a bigger trend: MCP isn't just for third-party tool integrations anymore, it's how AWS is wiring its own first-party agent tooling together.
Takeaway: If you're building a RAG or agentic-search pipeline and you're currently relying on prompt instructions to keep sources current and trustworthy, that's a control gap worth closing at the API layer, not the prompt layer — go check whether your search/retrieval tool already supports this kind of hard filtering, and if not, that's your next fix.