Supplementing, Not Reloading
When an internal knowledge base goes stale, the reflex is to re-crawl and regenerate embeddings — a full reload. In regulated industries, though, that pipeline carries approval and review gates that stretch from days to weeks, and the agent keeps answering from stale knowledge the entire time. The Web Search tool in Amazon Bedrock AgentCore, made GA in June 2026, opens a different route: ground on fresh web knowledge without touching your internal index, as a fully managed tool where data never leaves your AWS environment — zero data egress.
What Zero Egress Buys You in Regulated Settings
The tool combines Amazon's own search infrastructure — a proprietary web index plus a knowledge graph — and is exposed through an MCP connector on AgentCore Gateway. Each result returns ranked snippets alongside a source URL, title, and publication date, which is exactly the evidence recency judgments need. The same month raised Runtime quotas: InvokeAgentRuntime went from 25 to 200 TPS and active sessions from 1,000 to 5,000, so attaching web supplementation to high-traffic paths became viable too.
The Merge Rule: Web Assists, KB Wins
Bolting on managed web opens a new failure surface. Leave the priority between web results and the internal KB undefined and the agent blends them arbitrarily; ignore publication dates and it cites years-old documents as if they were current. Keep the design principle blunt: for facts the company is authoritative on — contracts, policies, rate schedules — the KB wins, and the web serves only to fill segments where the KB is empty or plainly outdated.
A Freshness-Supplement Architecture: Design Notes from Rollout to Operations
Set target numbers before anything else. A workable starting line: at least 70% of cited sources within the last N days (say, publication date under 180 days), 100% KB-wins application whenever a KB-web conflict is detected, web-search call rate under 30% of all queries, and web-search p95 latency under 2.5 seconds. Above all, make the "when do we call the web" trigger explicit — gate the call so it fires only when KB retrieval confidence falls below a threshold or the query carries a recency signal (this year, latest, recently amended, and the like).
Three failure patterns recur. First, an undefined priority lets a web snippet overwrite an internal rule — force a source-type tag (KB/web) at the merge step and hard-code KB adoption on conflict. Second, unparsed dates surface a document from years ago — treat the returned publication date as the primary key of the recency filter and downgrade confidence on results that carry no date. Third, unbounded web search drives cost and latency up together — cap it with the call gate, a result-count ceiling, and a per-session call cache.
Declare the recovery branches in advance. If web search times out or every result fails the recency filter, the agent answers from the KB alone but stamps the response with a "fresh web check failed" signal, and if the query carried a strong recency signal it escalates to a human-review path. When web and KB conflict and the KB is adopted, log that conflict pair and route it back as a KB-update candidate — the goal is not to abolish reloads but to let data set their priority.
The operations checklist starts with observability fields. Per query, record whether the web was called, the returned source URLs and dates, recency-filter pass/fail counts, KB-web conflict detections, and web-search latency — without these you cannot aggregate the targets above on a dashboard. Before release, run regressions against a scenario set built around recency-filter boundaries (N days ±1) and conflict cases, and verify that source URLs actually surface in responses and that the MCP connector holds least-privilege scope.
The improvement loop tracks two numbers each cycle. If the fresh-source ratio dips below the 70% target, separate whether the recency window N is too narrow or the web-call gate is clamped too tight; pull the top KB-web conflict types and feed them back as priorities for the reload queue. Run this loop and web grounding stops being a detour around the KB and settles into a buffer that refills freshness faster than the KB goes stale.
Adoption Decision Points at a Glance
The answer to staleness need not always be a reload. Supplement only the stale segments with zero-egress managed web, block outdated citations with a date-based recency filter, close conflicts with a KB-wins merge rule, and bound cost and latency with a call gate — declare those three axes in code and you can manage 70% fresh sources, 100% KB adoption on conflict, and a sub-30% web-call rate from one dashboard.
References
Amazon Bedrock AgentCore Web Search documentation — AWS
Release notes for Amazon Bedrock AgentCore (June 2026) — AWS