AWS Summit New York pushes Bedrock into managed RAG and grounded web search

AWS used Summit New York on 2026-06-17 to push Amazon Bedrock past its original framing as a managed catalog of frontier models. Two general availability launches landed in the same window: Amazon Bedrock Managed Knowledge Base, taking over the retrieval-augmented generation pipeline end to end, and Web Search on Amazon Bedrock AgentCore, the grounded search tool agents have been wiring up through Tavily and Perplexity until now. The pairing reads as AWS deciding that the layers between model and answer are platform surface area, not partner surface area.

What shipped

Amazon Bedrock Managed Knowledge Base reached GA with six native data connectors (Amazon S3, SharePoint, Confluence, Web Crawler, Google Drive, and OneDrive), Smart Parsing that handles content optimization without hand-tuned chunking, and an Agentic Retriever that resolves multi-hop queries across stored knowledge. The service takes over chunking, embedding refresh, and retrieval ranking, the three pieces most teams have been rebuilding for every project on top of Bedrock or a third-party vector database. Pricing meters by indexed data size and retrieval volume, and the service is available in US East, US West, Asia Pacific, and Europe regions at launch, with first-class integration into AgentCore Gateway and MCP-compatible agent frameworks.

Web Search on Amazon Bedrock AgentCore shipped the same day as a fully managed tool that lets agents reach the live web inside their reasoning loop. The implementation runs AWS-hosted search infrastructure plus structured knowledge-graph results, returns snippets, URLs, titles, and publication dates, and operates through the Model Context Protocol so agents written against the AgentCore runtime can call it without custom integration code. The pricing posture is unusually plain: no incremental tool charge beyond standard data transfer, with US East (N. Virginia) as the launch region and other geographies to follow. The strategic detail is that the tool keeps queries inside AWS rather than routing through Tavily, Perplexity, or Bing, closing one of the more common data-exfiltration paths in production agent stacks.

Bedrock AgentCore Harness also reached GA in the same announcement window, completing a Summit New York lineup that names AWS’s agent strategy as a layered platform (model, retrieval, tools, harness) rather than a single product. WAF Bot Control gained the ability to price and meter AI-bot access to content, Amazon Quick Autonomous Agents shipped for finance and CRM workflows, and S3 Annotations now attach up to 1 GB of queryable context per object.

Where this lands in the market

Managed RAG is the category where Vertex AI Search, Azure AI Search, Pinecone Assistant, and a long tail of integration vendors have been selling against teams who would rather not own the pipeline themselves. AWS shipping a GA service with first-party connectors and built-in agent orchestration narrows the differentiation those vendors can claim, and the connector list reads as a deliberate move on Microsoft (SharePoint, OneDrive) and Google (Google Drive) territory. Smart Parsing and the Agentic Retriever are the parts that signal AWS is trying to compete on retrieval quality rather than infrastructure plumbing.

The Web Search launch sits in a different lane. OpenAI’s hosted browser tool, Anthropic’s web search tool, and Perplexity’s Sonar API have defined what agent web grounding looks like, and AWS is the first hyperscaler to ship a comparable managed primitive that stays inside the customer’s cloud boundary. The MCP framing matters because it lets the same tool work for agents written against any MCP-compatible harness, not only AgentCore-native code. The clean cost story (no search-tool surcharge) appears to bet that compute and storage attachment is the real revenue lever, not per-query tool fees.

What’s worth watching

1. Retrieval quality against incumbents. Smart Parsing and the Agentic Retriever are testable claims. Independent benchmarks against Vertex AI Search and Pinecone Assistant will determine whether Bedrock Managed Knowledge Base wins on capability or only on procurement convenience. Anything less than parity makes the service a default for Bedrock-first shops and a non-event for everyone else.
2. Search-tool quotas as production tightens. Web Search shipped without an explicit per-query fee. As agentic workloads scale, watch for whether AWS introduces rate ceilings, freshness windows, or premium tiers; that meter shape will tell teams whether to treat the tool as a default primitive or as a fallback to a third-party search API.
3. Cross-cloud connector pressure. Native pulls from SharePoint, OneDrive, and Google Drive route enterprise content into AWS storage and retrieval before any model call. Whether Microsoft and Google respond with API restrictions, exclusivity terms, or competitive counter-moves will define how durable this connector lineup is over the next two quarters.

The plain reading is that AWS is no longer letting RAG, grounded search, and agent orchestration sit outside Bedrock. Each of the three GAs is in a category where AWS was previously content to let a partner or third-party service own the developer relationship; Summit New York names those categories as platform layers. Whether the platform displaces the incumbents depends on retrieval quality and connector durability holding up under production load.

Sources cited

1. AWS: Introducing Amazon Bedrock Managed Knowledge Base aws.amazon.com
2. AWS: Announcing Web Search on Amazon Bedrock AgentCore aws.amazon.com
3. AWS: Top announcements of the AWS Summit in New York, 2026 aws.amazon.com