AWS September Roundup: Bedrock, AgentCore and Strands Get Updates

AWS has released its monthly digest for AI developers for September 2026. It covers three products (Amazon Bedrock, Amazon Bedrock AgentCore and the Strands framework) and three main changes: a wider choice of models, faster serverless agents with built-in quality evaluation, and automatic knowledge base syncing with native enterprise connectors. The digest itself is short. It contains no benchmarks, latency figures or prices, so below we look at what is behind the wording and what you should verify yourself.

Three layers of one stack

The updates touch different levels of the platform. Bedrock handles access to models through a single API. AgentCore serves as the runtime for agents: it is where they are deployed, isolated and connected to tools. Strands works at the code level. It is an SDK in which the developer describes the logic of the agent itself.

The September set of updates addresses three typical problems when taking an agent to production. You need to pick a model for the task, run the agent without fiddling with infrastructure, and keep it supplied with current company data.

Models, agents, knowledge

The first change is a wider choice of models in Bedrock. For teams that do not want to be tied to a single vendor, this is a practical argument: you can swap the model without rewriting the integration. The digest does not say which models were added, so you will have to check the list in the Bedrock catalog.

The second change is more interesting. AWS talks about faster serverless agents with built-in evaluation. Agent quality evaluation has long been a bottleneck: teams write their own test suites and scripts to run scenarios. If AgentCore takes this on, the barrier to entering production drops. The caveat is that "faster" with no stated baseline remains marketing language. Cold start, response time and behavior under load need to be measured on your own scenarios.

The third change concerns knowledge bases. Automatic syncing with native enterprise connectors means the RAG index is updated as the sources change, rather than on a schedule that someone has to maintain. For companies with documents in SharePoint, Confluence and similar systems, this removes a routine chore. Which connectors are included also needs to be checked in the documentation.

What we know from the digest

• Period: September 2026, published on the AWS Machine Learning blog on October 9.
• Products: Amazon Bedrock, Amazon Bedrock AgentCore, Strands.
• Models: the choice in Bedrock has been expanded; the digest gives no list.
• Agents: serverless execution is faster, and built-in evaluation has been added.
• Knowledge: automatic knowledge base syncing and native enterprise connectors.
• Performance metrics and prices: not disclosed.

How to use this in practice

A team already running on Bedrock should do a short audit. First, compare the list of available models with your current configuration: there may be a cheaper or more accurate option for a specific task. Then run one of your existing agents through the built-in evaluation and compare the result with your own tests. If the numbers match, you can drop some of your homegrown tooling.

For those building RAG on corporate data, the main question is about connectors: do they cover your sources, how are access permissions handled, and how quickly do changes in a document reach the index? The last point determines whether the agent starts answering with stale data.

For those still choosing a platform, the digest is useful as a map of landmarks, but not as a basis for a decision. It is better to compare AgentCore with alternatives in a pilot under real load.