OpenAI’s agents can now run entirely inside Amazon’s cloud

Companies can now build OpenAI agents that run entirely inside Amazon Web Services, with their data never leaving AWS. OpenAI and Amazon launched Bedrock Managed Agents, powered by OpenAI, at OpenAI’s DevDay 2026 event in San Francisco on Tuesday. The product is in limited preview, says the AWS product page .

OpenAI built it with Amazon from the core of its Agents API. It added customisation so the agents work natively in AWS and with AWS resources, the company wrote in its DevDay recap .

Bedrock Managed Agents is built on OpenAI’s own agent framework, according to AWS. The managed runtime handles inference, memory and skills inside the customer’s environment. Each agent has its own identity and logs every action for auditing.

All inference runs on Amazon Bedrock, and customer data never leaves AWS, the product page says. AgentCore, Amazon’s platform for building and running agents with any model, is the default compute environment. AWS said AgentCore and Bedrock Managed Agents will add authorisation policy enforcement, agent discovery, observability and evaluation tools.

AWS lists Salesforce as a customer. Joe Inzerillo, Salesforce’s president of enterprise and AI technology, said customers can pair OpenAI’s models and AWS infrastructure with Salesforce data and controls.

“Data never leaves AWS, and security stops being the bottleneck between our customers and what comes next,” Inzerillo said in a statement on the AWS page.

In August, Salesforce put Anthropic’s Claude at the centre of its products. AWS said last week that almost 90% of Amazon’s early agent prototypes never shipped.

The Agents API that Bedrock Managed Agents builds on now supports computer use, OpenAI said. Developers can build agents that work through software in a browser OpenAI hosts. The API also brings in the multi-agent tools, tool search and context compaction used in Codex. OpenAI runs the underlying infrastructure.

Computer use is available through the API and in Codex and ChatGPT Work on Pro 500 and Enterprise plans. The user has to approve each new website the browser visits. That approval does not require confirmation before individual actions, OpenAI’s developer documentation says. Sign-in details go through a separate channel and stay out of the model’s input.

On Monday, OpenAI said its agents had posted users’ images online, affecting 53 ChatGPT users. At the same event, it launched its always-on dots agents , and said it is working with Microsoft to bring specialist dots to Agent 365. OpenAI also launched a $500 Pro plan and halved Pro 200 usage.