Amazon Web Services has launched a public preview of Amazon Bedrock Managed Agents powered by OpenAI, AWS’s Channy Yun wrote in the company’s weekly roundup on Oct. 5. The service is built on a customized version of OpenAI’s Agents API, adapted to run natively on AWS.
Agents built with it are optimized for OpenAI models and run entirely inside AWS, using the identities, permissions and governance controls customers already have, according to the roundup. Bedrock is AWS’s service for building with AI models from several providers.
How it works
AWS’s documentation describes the service as a way to run “stateful” agents, meaning they keep the conversation’s context across messages. A developer creates a session that names a model, instructions, tools, an access role and a place for the agent’s commands to run. The service manages the conversation and model calls.
AWS suggests it for multi-step jobs, such as investigating a codebase, processing documents or generating files. Agents can use skills, which are reusable instructions, and servers built on the Model Context Protocol (MCP), an open standard for connecting AI tools to data.
Customers choose where the agent’s tools run. They can use their own compute, such as a development machine or container, or Amazon Bedrock AgentCore Runtime, AWS’s managed option. The documentation says the connection between the two runs through OpenAI’s “codex exec-server” process, which makes an outbound link to the service.
The preview has limits. It’s available in three U.S. regions, and session input is text only, according to AWS’s preview page. Subagents aren’t supported, and there’s no built-in long-term memory.
AWS also warns that OpenAI’s own hosted Agents API and the Bedrock version “do not necessarily expose the same features or change at the same time.” Customers pay for model use and the AWS resources their application uses, and the documentation notes that a deployed example can keep running up charges when no agent is working.
Also in the roundup
AWS introduced Kiro workflows the same week. Kiro is AWS’s coding agent, and workflows let it carry out complex tasks from start to finish with multiple agents and less supervision. AWS says it has been building Kiro itself with workflows, including its new cloud sessions.
The roundup also announced a preview of the AWS Well-Architected Agent, which analyzes a customer’s AWS setup and recommends changes to improve cost, security, performance and resilience.
Bedrock added new frontier models as well, Yun wrote. They include OpenAI’s GPT-6.1 Sol, which OpenAI says comes close to GPT-6 Astra on demanding tests at about one-fifth of the cost. Anthropic’s Claude Sonnet 5.5 and SpaceXAI’s Grok 4.7 were also added, along with a faster “Ultrafast” tier for GPT-6 Astra that OpenAI says reaches up to 300 tokens per second.
Separately, AWS listed service changes. Amazon Mechanical Turk reached end of support as of Sept. 29, according to the roundup.
Source: AWS News Blog, Channy Yun, “AWS Weekly Roundup: Amazon Bedrock Managed Agents powered by OpenAI, Q3 service availability updates, Kiro workflows, and more (October 5, 2026)”
Additional sources: AWS Documentation, Amazon Web Services, “Amazon Bedrock Managed Agents, powered by OpenAI (preview)”; AWS Documentation, Amazon Web Services, “Preview availability and limitations”
