AWS

Amazon Bedrock AgentCore: Managed knowledge bases and web search for platform teams

Amazon Bedrock adds managed knowledge bases and integrated web search via AgentCore, letting teams combine private RAG sources and live web grounding inside AWS.

June 21, 2026·3 min read·AI researched · AI written · AI reviewed

AWS just made it trivial to stitch private knowledge and live web context into a single agent pipeline — and they did it by moving the indexing and search plumbing into Bedrock. That’s a huge operational win and an immediate new thing platform teams need to own.

The headline: Amazon Bedrock now offers a fully managed knowledge base with native connectors, multi-format parsing for spreadsheets/PDFs/HTML, and an agentic retriever. Those pieces are integrated into AgentCore so you can build retrieval-augmented agents without standing up your own indexing, storage, or orchestration stacks. Bedrock’s agent features also include integrated web search that can be configured to keep queries and results inside the customer's AWS boundary, enabling live grounding while limiting external egress.

Why this matters

Until now, RAG at scale forced teams into a slow loop: extract, normalize, index, maintain vector stores, and bolt on orchestration. Native connectors and multi-format parsing mean Bedrock will absorb common formats, normalize them, and keep them queryable. The agentic retriever is aimed at multi-step reasoning — think long-context chains that need compound retrieval strategies — and AgentCore ties the whole thing into Bedrock’s agent runtime.

The practical implication: teams can iterate on agent behaviors faster. The operational implication: you now have a managed service that embodies your retrieval policy, citation behavior, and access controls. Permissions, data retention, model grounding, and audit trails now live partly in the Bedrock runtime rather than solely in repos you control.

Integrated web search is consequential for compliance-minded orgs: mixing private KBs with live web facts while keeping traffic within the AWS boundary reduces some risk, but "inside AWS" is not the same as "safe by default." Search tools paired with an agentic retriever make it easier for models to act on external inputs; you still need explicit guardrails. The new guardrail capabilities help, but teams must instrument, monitor, and enforce them as code.

Other platform changes that matter

  • EC2 G7 is now generally available and is positioned for high-performance inference, graphics, and analytics workloads — a natural fit for low-latency model serving and visualization pipelines tied to Bedrock, SageMaker, or containerized stacks.

  • Amazon ECS added higher-resolution service metrics for autoscaling. These finer-grain signals let scaling policies react faster to spikes and tail latency, which can reduce overprovisioning for bursty microservices and shorten scale-up windows.

  • AWS also highlighted a set of AI-native operational capabilities — agent-driven DevOps and security automations and continued SageMaker inference optimizations — signaling a push toward agent-first operations: automations that can take action across pipelines.

What platform teams must do (opinionated)

This is the right call from AWS: abstracting the boring parts of RAG lets teams focus on semantics and guardrails instead of custom indexing glue. But it also shifts where risk accumulates. If you treat Bedrock-managed KBs and AgentCore as just another managed datastore without adding runtime auth, telemetry, and threat modeling, you’ll be the team that discovers an agent exercised permissions you didn’t expect.

Start with three realities:

  1. Treat managed KBs as first-class data planes: implement RBAC, data-class tagging, and audit logging.
  2. Instrument agent decisions: log retrieval traces, citations, and the sequence of tool calls so post-mortems are possible.
  3. Bake guardrails into CI/CD and the service mesh: agents should be feature-flagged, canaryed, and observable like any other service.

One neat follow-up: if you want a deeper look at how Bedrock packaged these features, I wrote a closer take on the managed KB and AgentCore combo Amazon Bedrock AgentCore: Managed Knowledge Base and Web Search for Platform Teams.

Parting thought

AWS is making the obvious trade: reduce developer friction by owning retrieval and web grounding, and in exchange surface new runtime responsibilities for platform teams. Expect faster experiments and more powerful agents — but also more incidents from forgotten permissions and ungated tool invocations. The next battleground for platform engineering isn’t raw throughput or cheaper GPUs; it’s runtime governance for autonomous systems. Platforms that get that right will win.

Sources

amazon-bedrockagentcoreec2-g7ecs-auto-scaling
← All articles
AWS

Amazon Bedrock Managed Knowledge Bases: connectors, Smart Parsing, and agent retrievers for platform teams

Amazon Bedrock now adds Managed Knowledge Bases with connectors, Smart Parsing, and agent retrievers, moving RAG plumbing into a managed retrieval plane.

Jun 24, 2026·3mamazon-bedrockbedrock-agentcore
AWS

Amazon Bedrock Agent Core Web Search: Agents Can Now Ground Responses in Live Web Content

Bedrock's Agent Core adds Web Search so agents can cite live web content without you running a search index, introducing new operational and security risks.

Jun 22, 2026·3mamazon-bedrockagentcore
AWS

Amazon Bedrock AgentCore: Managed Knowledge Base and Web Search for Platform Teams

Bedrock's AgentCore adds a managed knowledge base and web search with VPC-friendly connectors, reducing vector-store ops but increasing audit and IAM scope.

Jun 20, 2026·3mamazon-bedrockagentcore