The AWS Machine Learning Blog post “ICYMI: What landed for AI builders in September 2026” was published on October 9, 2026, and recaps September updates across Amazon Bedrock, Amazon Bedrock AgentCore, and Strands. AWS presents the updates as changes to model choice, agent operations, and knowledge-base management. Reading them together helps show the range of work covered: some announcements add model options, others describe how agents run, and another set concerns scheduled refreshes of data used by knowledge bases.

This is a recap published by AWS, so its statements about availability and product characteristics are the organization’s own descriptions. They do not by themselves establish that every feature fits every deployment. The practical relevance depends on the models, agent architecture, and data sources a team uses. The September summary nevertheless gives a concrete view of several areas in which the platform was extended.

Managed agents and runtime operations

AWS says Amazon Bedrock Managed Agents, powered by OpenAI, is available in public preview. The post says it supports OpenAI models while keeping data within AWS. For teams building agent applications, these are two relevant parts of the announcement: model support and AWS’s description of where data remains. Because the offering is in public preview, the statement describes availability at that stage rather than a final product status.

The post also describes a newer AgentCore runtime. AWS says it has more efficient memory management and lower cold-start latency for serverless agents. It also mentions pay-as-you-go pricing and sessions that scale to zero when idle. These details describe aspects of the operating profile: startup latency, resource use, and idle behavior can all matter when teams plan how an agent runs and how its use is billed.

Each point refers to a different part of operations. Lower cold-start latency concerns the moment an agent starts after being idle. More efficient memory management concerns runtime resource use. Scaling sessions to zero and paying according to usage describe the capacity and billing model reported by AWS. It is therefore useful to read the announcement as a set of related characteristics rather than reducing it to a single performance metric: the recap links runtime behavior with what happens when a session is inactive.

Strands tools for building and deciding

For Strands, AWS describes an open-source agent harness that, according to the post, matches popular harnesses on accuracy while using 28 percent fewer tokens. The comparison combines two different criteria: quality as measured by the stated accuracy comparison and the amount of tokens used. The post presents the harness as an option for developers who want to consider both an agent’s results and its token consumption.

Another item is Strands Decider 2B, an open-source decision model with two billion parameters. AWS describes it as selecting between predefined options and delivering answers locally in approximately 115 milliseconds. The model’s stated role is important context: it chooses from options that have already been defined. The time is associated with local answers and is presented as an approximate figure, not as a universal guarantee for every environment.

Together, the two items concern different parts of agent work. A harness relates to assembling and running agent logic, while Decider 2B handles selection among options. The figures in the announcement can inform an initial review, but they do not by themselves determine how a solution will behave with a particular input or configuration. AWS includes both under its September Strands updates.

More model choice on Bedrock

AWS says its Astra, Sol, and Luna models are generally available on Amazon Bedrock. For GPT-6 Astra, the post cites support for up to 1 million input tokens. GPT-6 Astra Ultrafast is described as offering up to 6 times faster inference in the API, at up to 300 tokens per second. The announcement thus pairs a general-availability statement with particular parameters related to input capacity and inference speed.

The post also lists Kimi K3 as available on Amazon Bedrock. According to AWS, it has a 1-million-token context window, native vision, and built-in prompt caching. These are three distinct characteristics: context size, handling of visual input, and caching prompts for reuse. For teams comparing available models, the recap gives these as reference attributes rather than recommending one model for every use case.

This expansion of choice may matter to teams considering different models within the same platform. The recap itself does not provide independent test results or every detail needed for a specific scenario. The listed capacities and speeds should therefore be read as what AWS announced in the post. A monthly recap records new options; it does not replace a team’s own evaluation of models for its tasks.

Scheduled knowledge-base refreshes

For knowledge bases, AWS says Amazon Bedrock Managed Knowledge Base supports automatic refresh scheduling for all native data-source connectors. Schedules can be daily, weekly, or monthly. For people managing data sources, the key point is that the post describes scheduled refreshes and names three intervals. This gives teams a way to align a refresh schedule with how often source information changes, where the chosen schedule and connector permit it.

AWS also describes new native connectors for ServiceNow, Confluence Data Center, Salesforce, and Zendesk. In this context, a connector links a knowledge base to a source from which information can be drawn. The announcement adds four named enterprise systems to the available options. Alongside scheduled refreshes, this highlights two practical parts of knowledge-base management in the post: connecting a source and setting a regular synchronization cadence.

It helps to distinguish the scope of the announcement from assumptions about a particular deployment. AWS refers to scheduled refreshes for all native data-source connectors and names the connectors that were added. The post does not itself describe the configuration of a specific source or the resulting quality of a knowledge base. It provides information about available capabilities that a team can consider in relation to its own sources and refresh needs.

How to read the September recap

As AWS summarizes them, the September updates span three layers of AI application development: model availability and choice, agent tools and operations, and knowledge-base management. Bedrock features announcements about the general availability of OpenAI models and Kimi K3, as well as scheduled refreshes and additional connectors. AgentCore is associated with memory, cold starts, and scaling idle sessions. Strands includes a harness and a decision model with different stated characteristics.

A practical way for readers to use the recap is to identify which announcements relate to their own work. A team building agents may focus on the runtime and Strands tools; a team selecting a model may review availability and the parameters listed; a team maintaining knowledge sources may consider refresh schedules and new connectors. These points all come from one AWS post, published October 9, 2026, that recaps updates from September.