AWS September update expands models, agent runtime and evaluation tools

AWS has published a consolidated account of its September updates for AI builders, spanning Amazon Bedrock, AgentCore and the open-source Strands ecosystem. The changes target three production concerns: model choice, efficient agent operation and governed access to current enterprise information.

The update is more than a model catalogue. It shows AWS assembling a layered agent platform where developers can choose a model, run a durable session, evaluate behaviour, connect business data and apply existing cloud identity and audit controls.

Managed Agents adds an OpenAI path

Amazon Bedrock Managed Agents powered by OpenAI entered public preview. AWS says customers can build agents around OpenAI models while retaining IAM permissions, CloudTrail auditability, durable sessions and built-in human approval workflows.

This arrangement can reduce integration work for organisations already governed through AWS. Preview users should still test model-specific data handling, regional availability and tool behaviour, and should separate approval for a plan from approval for each consequential action.

AgentCore runtime becomes more elastic

The updated AgentCore runtime improves memory management and lowers cold-start latency for serverless agents. Sessions can scale to zero when idle and run in hardware-isolated environments, with pay-as-you-go pricing rather than pre-provisioned peak capacity.

Serverless economics suit uneven demand, but production teams need to measure wake-up delay, long-running session behaviour and concurrency limits. A faster average start does not remove tail latency or the need to protect state during retries.

Strands focuses on efficiency and decisions

A new Strands harness is reported to match popular agent harnesses on accuracy while using 28 per cent fewer tokens. It includes context management, prompt caching and memory and can be deployed outside AWS.

Strands Decider 2B takes a different approach: it is a small open model that chooses among predefined options rather than generating prose. AWS reports roughly 115-millisecond local responses for tasks such as routing, tool selection and guardrails. Restricting the output space can make fast operational decisions easier to validate.

Bedrock broadens frontier-model choice

The September portfolio includes OpenAI Astra, Sol and Luna families for different intelligence, speed and volume requirements. AWS also highlights recent Claude releases, Moonshot AI’s Kimi K3 and xAI’s Grok 4.6 and 4.7.

A larger catalogue supports workload routing, but it also increases evaluation and governance work. Teams should maintain a model registry showing approved tasks, regions, prices, context limits, retention settings and fallback behaviour. A model swap is only simple when those differences are explicit.

Evaluation moves closer to the runtime

A production agent needs evidence about task success, safety and cost, not only benchmark scores for its underlying model. AWS is positioning managed evaluation and observability alongside the runtime so developers can assess complete workflows.

Useful evaluations should include tool selection, permission boundaries, recovery from partial failure and the quality of human hand-offs. They should also record the prompt, model and tool versions required to reproduce a result.

Enterprise connectors keep knowledge current

The September releases include more direct ways to synchronise knowledge bases with business systems. Native connectors can reduce custom ingestion code and help agents retrieve up-to-date material instead of relying on stale exports.

Synchronisation must preserve source permissions and deletion. An agent should not expose a document merely because it reached a common index. Administrators need evidence that access is checked at retrieval time and that revoked content is removed promptly.

System design matters more than one model

AWS frames the update around the infrastructure surrounding an agent: context, actions, security, measurement and data governance. This reflects the shift from experiments to systems expected to work across long tasks and sensitive enterprise applications.

Builders should resist treating the full platform as a default architecture. Each layer adds capability and operational surface. A small decision model, managed agent or knowledge connector should be included because it solves a measured problem, not simply because it is available.

September’s releases form a production stack

Taken together, the updates give AWS customers more ways to assemble production agents without committing to one model or framework. Serverless runtime, open tools, model breadth and governed data access can shorten the path from prototype to deployment.

The opportunity comes with a requirement for disciplined routing and evaluation. Choice is valuable only when teams know why one model or runtime is selected and can audit the result. AWS’s September release set provides more building blocks; production quality still depends on how deliberately they are combined.

Cost management should sit alongside quality testing. Multiple models, caches, serverless sessions and retrieval calls can make a single user request difficult to price. End-to-end tracing should attribute each charge to the task and expose when retries or oversized context dominate spending.

Security teams should review the combined authority of connectors and tools. A model with restricted behaviour may still retrieve sensitive material or trigger a powerful action through the surrounding agent. IAM and human approval are most effective when permissions are narrow, temporary and connected to a clearly identified workload.