AWS previews Bedrock agents built with OpenAI, under AWS controls

AWS has opened a preview of Bedrock Managed Agents, a joint AWS–OpenAI service for building agents that run inside AWS. The draw is the combination of OpenAI-powered agent capabilities with AWS identity, permissions and governance controls already used by customers.

AWS AI Watch analysis

What happened

AWS says Bedrock Managed Agents is built on a customised version of OpenAI’s Agents API and is available in three US regions: Northern Virginia, Oregon and Ohio. It manages agent state, tool use, code execution and work across multiple steps.

The preview supports durable sessions, reusable skills and connections to tools, including through Model Context Protocol servers. Each agent gets its own IAM role; customers can require human approval before consequential actions and record supported API activity in CloudTrail. AWS says there is no additional charge for the service during preview, though customers still pay for the AWS resources their agents use. Pricing may change at general availability.

Our top picks

• Agents run inside AWS
  AWS says the service is integrated with its resources and uses customers’ existing identities, permissions and governance controls.
• Durable sessions
  Agents can retain messages, tool calls and intermediate results so work can resume later with new information.
• Reusable skills and MCP tools
  Developers can add specialised procedures and connect agents to external tools through MCP servers.
• An IAM role for each agent
  Separate identities give operators a way to define what each agent can access.
• Approval before consequential actions
  Customers can put a human checkpoint in the path of actions that matter.
• CloudTrail records supported API activity
  Operators can use AWS’s audit tooling to track supported activity by agents.

Why it matters

This gives AWS customers a way to build agents around OpenAI’s technology without moving the agent runtime outside AWS. For organisations already invested in AWS, the integration of IAM, approval controls and audit logging may matter as much as the model: useful agents need to reach tools, but operators need to know what they can do and how to trace their activity.

The service is still a preview, initially limited to three US regions. And “no additional charge” is not the same as free: the underlying AWS resources still run up a bill.

Our read

The interesting proposition is not simply OpenAI inside Bedrock. It is an attempt to make multi-step agents fit into the permissions and oversight machinery companies already use. That could lower the friction of trying agents in AWS environments, while giving operators more familiar control points than a standalone experiment usually offers.

The practical test will be whether those controls are clear and dependable in real workflows, not just available as feature names. Teams considering the preview should check regional availability, resource costs and which activity is covered by CloudTrail before handing an agent the keys. Preferably not the master key.

What to watch

• Whether AWS expands the preview beyond its three initial US regions.
• What pricing and availability look like when the service reaches general availability.
• How much agent activity is covered by CloudTrail and how approval controls work in practice.

Discussion spark: Would you trust an AI agent with access to your AWS tools if it had its own IAM role and human approval for consequential actions, or would you still keep it out of production?

not affiliated with or endorsed by Amazon Web Services (AWS)