AWS has added new agent controls, runtime changes and a broader range of models to Bedrock and its developer tools. The practical shift is towards giving builders more ways to manage cost, permissions and the systems their agents can reach, alongside more choice about which models to run.
AWS AI Watch analysis
What happened
In a recap published on 9 October, AWS outlined September updates across Amazon Bedrock, Bedrock AgentCore and Strands. Its account describes several changes with direct consequences for people building AI applications:
Our top picks
- OpenAI-powered managed agents in Bedrock
Now in public preview, they include durable sessions, human approval workflows, IAM permissions and CloudTrail auditability, AWS says. - A leaner AgentCore runtime
AWS says improved memory management lowers cold-start latency, while serverless agents can scale to zero when idle and use pay-as-you-go pricing. - A new Strands harness
AWS says the open-source toolkit matches popular harnesses on accuracy while using 28 per cent fewer tokens. That is a vendor-reported comparison, not an independent benchmark. - A small decision model for agent tasks
Strands Decider 2B is an open-source, two-billion-parameter model designed to choose between predefined options, such as tool selection and routing. AWS says it can respond locally in about 115 milliseconds. - More model and data-source options
The recap lists OpenAI, Anthropic, Moonshot AI and xAI models on Bedrock, and new native knowledge-base connectors for ServiceNow, Confluence Data Center, Salesforce and Zendesk.
Why it matters
The changes address some unglamorous but important agent-building questions: what an agent can access, whether a person can approve consequential actions, how quickly an agent starts, and how much it costs while idle. The new connectors could also reduce the custom ingestion code teams need to maintain as their business information changes.
More model choice is useful only if teams can compare models against their own workloads and budgets. AWS’s recap names options and features, but it does not provide independent evidence that each one will be the best fit for a particular job.
Our read
The strongest part of this bundle is the attention to the machinery around agents, not just the latest model on the menu. Human approvals and auditable permissions are concrete controls; scale-to-zero is a potentially useful way to avoid paying for idle capacity. The 28 per cent token claim is worth testing against real workloads before treating it as a saving in the bank.
What to watch
- When the managed-agent preview becomes generally available, and on what terms.
- Whether AWS publishes further detail on the Strands harness comparison.
- How the new connectors handle permissions and updates to source data.
Discussion spark: When deploying business AI agents, which should come first: tighter human approval and audit controls, or lower operating costs?
Sources and evidence
- ICYMI: What landed for AI builders in September 2026 (9 October 2026, 15:38 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)