On 29 September 2026, during OpenAI’s DevDay in San Francisco, Amazon Web Services and OpenAI announced Bedrock Managed Agents, a limited-preview product that lets companies build OpenAI agents that run entirely inside AWS. The Next Web framed it as a direct answer to the enterprise security objection: customer data never leaves Amazon’s cloud.
Per the AWS product page cited in coverage, 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; AgentCore, Amazon’s platform for building agents with any model, is the default compute environment.
OpenAI said it built the offer with Amazon from the core of its Agents API, with customisation so agents work natively with AWS resources. AWS said AgentCore and Bedrock Managed Agents will add authorisation-policy enforcement, agent discovery, observability and evaluation tools.
Salesforce is listed as a customer. Joe Inzerillo, Salesforce’s president of enterprise and AI technology, said customers can pair OpenAI models and AWS infrastructure with Salesforce data and controls: “Data never leaves AWS, and security stops being the bottleneck.” In August Salesforce put Anthropic’s Claude at the centre of several products; both labs coexisting in the AWS/Salesforce ecosystem illustrates a multi-model strategy.
The Agents API underneath now supports computer use: agents that operate software in an OpenAI-hosted browser. The user must approve each new website; confirmation is not required before every individual action. Sign-in credentials travel on a separate channel, out of the model’s input.
AWS said last week that almost 90% of its early agent prototypes never shipped. Bedrock Managed Agents aims to shorten that funnel by offering identity, logs and data residency as defaults, not as a bespoke project for every platform team.
The launch sits beside dots and the White House pact. For a CIO, the practical question differs: prefer a ChatGPT dot with Guardian, or a Bedrock agent whose traffic and storage sit under their AWS account and IAM controls? The second fits regulated security reviews (finance, health, public sector) more cleanly.
Residual risks: computer use still depends on site approvals a user may grant too broadly; auditable logs alone do not stop harmful actions if the tool policy is wide; and limited preview means unstable APIs and quotas.
Competitively, Microsoft pushes Agent 365 and dots integration; Google offers agents on Vertex; Amazon answers with strict Bedrock residency. OpenAI becomes the model engine across several clouds at once—a distribution strategy that reduces single-hyperscaler dependence and complicates any “one stack” narrative.
EU compliance implications: if inference and data stay in European AWS regions, the transfer argument simplifies, but the model provider (OpenAI) remains GPAI under the AI Act. Residency does not erase documentation or serious-incident duties.
Neural Pulse suggests enterprise teams trial the preview on a low-risk case (internal ticket summarisation) measuring latency, log fidelity, Bedrock cost and computer-use approval friction—only then expand to actions with external effects.
The close: Bedrock Managed Agents does not solve GPT-6 Astra alignment; it solves where data lives while the agent thinks. In a September obsessed with sandbox escapes, “never leaves AWS” is a sales slogan—and a real control—that many boards will grasp faster than a moral pact on the White House lawn.
