OpenAI introduces Stateful Runtime for agents in Amazon Bedrock
Original: Introducing the Stateful Runtime Environment for Agents in Amazon Bedrock View original →
What the new runtime does
On February 27, 2026, OpenAI said it is working with Amazon on a Stateful Runtime Environment for Agents that runs natively in Amazon Bedrock. The company framed the launch around a practical problem: AI agents may be good at reasoning, but production systems break down when teams have to manage long-running workflows, state, approvals, tool execution, and error recovery around stateless APIs.
According to OpenAI, the new runtime is designed to keep a working context across many steps instead of forcing developers to stitch together disconnected requests. OpenAI said the runtime carries forward memory and history, tool and workflow state, environment use, and identity and permission boundaries. It also said the system runs inside the customer’s AWS environment and is optimized to work with AWS services, which is meant to make governance and security easier for enterprise teams.
How it fits into the OpenAI-Amazon partnership
The runtime is one piece of a broader OpenAI-Amazon deal announced the same day. Amazon said AWS will be the exclusive third-party cloud distribution provider for OpenAI Frontier, the company’s enterprise platform for building and managing teams of AI agents. Amazon also said OpenAI plans to consume about 2 gigawatts of Trainium capacity through AWS infrastructure to support Stateful Runtime, Frontier, and other advanced workloads, with the runtime expected to launch in the next few months.
That larger context matters because it shows this is not just a feature announcement. OpenAI and Amazon are trying to define a managed execution environment for long-horizon agent work, one that sits closer to the customer’s cloud boundary instead of leaving orchestration entirely to application teams. The move also indicates that control over the agent runtime layer is becoming strategically important, separate from control over the model itself.
Why developers and platform teams should care
If the rollout works as described, the immediate benefit is less scaffolding. Teams building multi-system customer support, internal IT automation, finance approvals, or sales operations workflows would spend less time building custom state handling and more time on business logic. In practice, that can be the difference between an impressive demo and an auditable production workflow.
The deeper significance is market structure. The agent stack is starting to separate into at least two layers: raw model endpoints and managed execution environments. OpenAI’s Bedrock runtime suggests that cloud providers and model vendors now see the control plane around agents, not just the model weights or API, as a critical part of the competitive stack.
Sources: OpenAI, Amazon, OpenAI and Microsoft
Related Articles
This is a distribution story, not just a usage milestone. OpenAI says Codex grew from more than 3 million weekly developers in early April to more than 4 million two weeks later, and it is pairing that demand with Codex Labs plus seven global systems integrators to turn pilots into production rollouts.
The bottleneck moved from GPUs to the API layer, and OpenAI changed the transport to keep up. By adding WebSocket mode and connection-scoped caching to the Responses API, the company says agentic workflows improved by up to 40% end-to-end and GPT-5.3-Codex-Spark reached 1,000 tokens per second with bursts up to 4,000.
OpenAI Developers said recent Codex usage data suggests developers are handing off long-running work like refactors and architecture planning at the end of the day. In a follow-up reply, the account said tasks started at 11 pm are 60% more likely than other tasks to run for 3+ hours.
Comments (0)
No comments yet. Be the first to comment!