AWS has recapped September 2026 updates to Amazon Bedrock, Bedrock AgentCore and Strands that expand how enterprise teams can build, run and ground AI agents. For builders, the changes add managed-agent controls, serverless operations, model-selection options and native ways to keep connected business data current.
Run faster agents and evaluate what matters
Amazon Bedrock Managed Agents, powered by OpenAI, is available in public preview. AWS says the service lets customers use OpenAI models while retaining data in AWS, reusing IAM permissions and keeping audit records in CloudTrail. Durable sessions and human-approval workflows are intended to reduce development work and add controls to agent operations.
AWS also updated the AgentCore runtime for serverless agents. The company says improved memory management and lower cold-start latency mean customers pay for actual usage rather than peak memory, without pre-provisioning capacity. Sessions scale to zero when idle and run in hardware-isolated environments; AWS does not provide a numerical cold-start comparison in this roundup.
On the open-source side, Strands harness packages context management, prompt caching and memory into an agent harness that can be started from one line of Python or TypeScript, according to AWS. Strands claims it matches popular harnesses on accuracy while using 28% fewer tokens. Strands Decider 2B is a 2-billion-parameter decision model that chooses among predefined options rather than generating text; AWS says it returns local answers in about 115 milliseconds for tool selection, routing and guardrails.
Find the right model for each workload
OpenAI’s Astra, Sol and Luna models are now generally available on Bedrock. AWS positions GPT-6 Astra for complex decisions, document analysis and software development, with up to 1 million input tokens. Its Astra Ultrafast tier is claimed to deliver up to six times faster API inference and up to 300 tokens per second. GPT-6.1 Sol and GPT-6 Sol target recurring coding and professional work, while GPT 6.1 Luna is aimed at high-volume extraction, summarization, classification and routing.
The expanded catalog also includes Claude Fable 5.1, Claude Opus 5.5 and Claude Sonnet 5.5. AWS says Opus 5.5 uses adaptive thinking and offers an effort setting that caps reasoning depth; it says Sonnet 5.5 has 30% lower cost per task and 30% faster speed than Claude Sonnet 5. AWS also highlights Kimi K3 and Grok 4.6 and 4.7: Moonshot AI’s claims for Kimi K3 include 2.8 trillion parameters and 2.5-times-faster applications, while the Grok models have a 500K-token context window plus configurable reasoning effort and self-verification.
Get more control over your enterprise AI agents
Amazon Bedrock Managed Knowledge Base now supports daily, weekly and monthly automatic refresh schedules for every native data-source connector. A user-managed setup for SharePoint, OneDrive and Confluence can reduce reliance on administrator-managed service accounts when users already have access. New native connectors for ServiceNow, Confluence Data Center, Salesforce and Zendesk automate crawling, metadata extraction and incremental synchronization, reducing custom ingestion code for support content, internal documentation and operational knowledge.
Get started
AWS points builders to its Bedrock samples, the AgentCore CLI and the Strands Harness SDK for implementation. Source: AWS Machine Learning Blog
Definition. Amazon Bedrock is AWS’s platform for building AI applications and agents with model access, operational controls and connected enterprise knowledge.
| Update area | What AWS says changed |
|---|---|
| Bedrock Managed Agents | Public-preview managed agents powered by OpenAI with AWS data retention, IAM reuse, CloudTrail records, durable sessions and human-approval workflows. |
| AgentCore runtime | Improved memory management, lower cold-start latency, usage-based memory billing, scale-to-zero sessions and hardware-isolated environments. |
| Strands | A harness with context management, prompt caching and memory for Python or TypeScript. |
| Managed Knowledge Base | Automatic refresh schedules and native connectors for ServiceNow, Confluence Data Center, Salesforce and Zendesk. |
Key takeaways
- Amazon Bedrock Managed Agents is in public preview and is powered by OpenAI.
- AgentCore runtime updates include improved memory management, lower cold-start latency and scale-to-zero sessions.
- Strands harness packages context management, prompt caching and memory for Python or TypeScript agent development.
- OpenAI’s Astra, Sol and Luna models are generally available on Bedrock.
- Managed Knowledge Base supports daily, weekly and monthly refresh schedules for native data-source connectors.
- New native connectors cover ServiceNow, Confluence Data Center, Salesforce and Zendesk.
FAQ
What is Amazon Bedrock Managed Agents?
It is a public-preview service powered by OpenAI that AWS says lets customers use OpenAI models while retaining data in AWS, reusing IAM permissions and keeping CloudTrail audit records.
What changed in the AgentCore runtime?
AWS says its serverless-agent runtime now has improved memory management and lower cold-start latency, bills for actual usage rather than peak memory, and scales idle sessions to zero.
Which OpenAI models are generally available on Bedrock?
AWS says OpenAI’s Astra, Sol and Luna models are generally available on Bedrock.
How can Managed Knowledge Base keep enterprise data current?
It now supports daily, weekly and monthly automatic refresh schedules for every native data-source connector.
