AWS has opened Amazon Bedrock Managed Agents, powered by OpenAI, to public preview. Jointly developed by AWS and OpenAI, the service adapts OpenAI’s Agents API to run stateful, multi-step agents entirely within AWS environments.
Bedrock Managed Agents preserves session state, selects tools, executes code, and carries work across multiple decisions. Developers can add reusable skills and MCP servers, run tools on self-hosted compute or Amazon Bedrock AgentCore Runtime, and resume durable sessions containing messages, tool calls, and intermediate results.
AWS governance is built into the runtime. Each agent receives an IAM role, consequential actions can require human approval, and supported API activity is recorded through CloudTrail.
The preview is available through bedrock-mantle endpoints in US East (N. Virginia), US West (Oregon), and US East (Ohio). AWS currently charges no additional fee for Bedrock Managed Agents beyond the model inference and AWS infrastructure the application consumes.
Important preview limitations include text-only input and no documented support for subagents, programmatic tool calling or code mode, cross-region inference profiles, or integrated long-term memory. The service-managed conversation and files in the customer’s execution environment also have separate lifecycles.
Why it matters
This is a significant distribution step for production AI agents. AWS customers can now combine OpenAI’s agent harness with their existing identities, permissions, compute, logging, and governance controls instead of assembling those operational layers independently.
The runtime choice also connects the preview to AWS’s broader AgentCore execution architecture, while preserving a self-hosted path for teams that need their own workspace, network access, or compute environment.
For builders, it reinforces a broader shift: the competitive agent stack increasingly includes durable state, execution environments, reusable skills, tool protocols, approvals, and auditability—not only the underlying model. Because this is a public preview, teams should test compatibility and cost behavior before relying on it for production workloads.