OpenAI Agents Move Into AWS: Seven Years of Microsoft Exclusivity End as Sovereign Enterprise Agents Arrive

What happened: OpenAI and AWS teamed up to lock agents inside your own cloud

Announced at OpenAI's DevDay on September 29 and confirmed by the AWS News Blog on October 5: Amazon Bedrock Managed Agents is in public preview. This is OpenAI's Agents API deeply customized into an AWS-native version — model inference runs on Bedrock, while agent sessions, memory, tool calls, and multi-step orchestration all live inside the customer's own AWS account.

The key design: every agent gets its own AWS identity (IAM role), explicit permission boundaries, and complete operation logs. Two execution environments: bring your own compute, or the managed Bedrock AgentCore Runtime. Reusable skills, MCP server connections, and human-approval gates before critical actions are all supported. No extra charges during preview beyond underlying resource consumption.

Why now: enterprise agents have been stuck at security review for two years

The enterprise agent playbook of the last two years is depressingly consistent: a dazzling POC, then a hard stop at security review. Legal won't sign — prompts, memory, and tool output leave the corporate account. Procurement won't approve — there's already a big AWS contract, and nobody wants a second direct pipe to OpenAI. The capability problem was solved long ago; what always blocked agents was 'where does the data flow'.

Bedrock Managed Agents sells a closed door: inference on Bedrock, runtime inside the customer environment, data never leaving AWS. In the words of Salesforce's president of enterprise and AI technology: 'Data never leaves AWS, and security stops being the bottleneck between our customers and what comes next.' For seven years, production OpenAI effectively meant Microsoft's cloud — that rule just broke, and the cloud neutrality war is officially on.

Technical notes: Codex harness plus durable sessions

Two technical details worth noting. First, the agent core uses OpenAI's Codex harness — the same execution framework driving Codex, not a watered-down version. Second, sessions are durable: messages, tool calls, and intermediate results persist, pausable and resumable. That matters enormously for long-running agents; state management was the biggest pitfall in most homegrown setups.

The new models landing on Bedrock alongside deserve a look: GPT-6.1 Sol targets agentic coding, reportedly approaching GPT-6 Astra levels at roughly one-fifth the cost; GPT-6 Astra gets an UltraFast tier claiming 6x faster API inference; Claude Sonnet 5.5 joined the shelf too. AWS clearly wants Bedrock to be the one-stop shop for 'good models plus good governance.'

Our take: vibe coders won't use it soon, but should read the wind

Honestly, this product isn't built for indie developers — its customers are banks, retailers, and telcos stuck on compliance. But every vibe coder should read the wind: 'sovereignty' is becoming standard vocabulary for agent infrastructure. IBM's self-hosted Bob, YTL's ILMUcode keeping data inside Malaysia, now AWS keeping data inside the customer account — three big stories in one quarter, all telling the same story.

The practical lesson for indies: if your vibe product ever wants enterprise customers, 'where data flows' will block the deal earlier than 'how flashy the features are.' Architect for private deployability now — config separated from code, bring-your-own-model endpoints, swappable data layers — so you have an answer when someone asks.