OpenAI's DevDay on September 29 was sold on agents that never log off. The more consequential change sat underneath the headline. In one week, OpenAI, Amazon, Red Hat and Nvidia all answered the same question: where does an AI agent actually run? An agent is no longer a feature you call. It is a workload you host, and that turns AI adoption from a software purchase into an infrastructure decision.
What Actually Happened
According to Decrypt's account of the event, OpenAI announced more than 20 launches. The headliner was Dots: always-on agents powered by GPT-6 Astra, each with its own cloud computer and web browser, connected to more than 4,000 apps and reachable through ChatGPT, Slack or Microsoft Teams. The first dot is included with Pro and Business Premium plans in eligible markets. OpenAI also released GPT-6.1 Sol, which it says nearly matches Astra on agentic coding and computer use at one-fifth of the token price, and an Ultrafast tier that runs up to six times faster at up to six times the standard rate.
Two infrastructure announcements received less attention. Bedrock Managed Agents, built with Amazon, runs OpenAI agents entirely inside AWS. Private Intelligence pairs zero data retention with automated safety checks, with a Private Inference preview on confidential-computing hardware due this autumn.
On September 30, the OpenClaw Foundation released OpenClaw Enterprise, a free, open-source control plane for persistent agents. According to Techzine, the project began inside OpenAI, was donated to the foundation, and is being developed with Red Hat and Nvidia. It adds multi-tenancy, hard security boundaries, sandboxing and audit trails, runs on Kubernetes, and lets organisations swap out the model, harness or sandbox. The foundation says many IT departments currently ban agent platforms outright because no shared security standard exists. Quartz reports the platform is pre-1.0 and recommended only for internal pilots, with a 1.0 release planned later in 2026.
The Quiet Shift: Agents Now Need an Address
For three years, enterprise AI decisions were model decisions. An application sent a request to an API, received an answer, and the transaction ended. A persistent agent breaks that pattern. It holds access to systems, keeps memory between tasks, acts on a schedule and keeps working after staff go home. It behaves like an employee and a server at once, so it needs what both need: a location, an identity, permissions and someone accountable for it.
This week's announcements lay out the three available answers. Vendor-hosted: OpenAI runs the dot's computer and you rent the outcome. Your cloud: Bedrock Managed Agents puts the same capability inside an AWS account you already govern. Your own infrastructure: OpenClaw Enterprise runs on Kubernetes clusters you operate. These are the same three options businesses faced for email and ERP a decade ago, and the trade-offs are familiar. Speed of deployment runs in one direction; control over data and cost predictability runs in the other.
The second-order effect is on the budget. The price of intelligence is falling quickly, with Sol offering near-flagship capability at a fifth of the cost. The cost of operating agents is moving elsewhere: compute time for agents that run continuously, integration work to connect them to business systems, and premium tiers for speed. An always-on agent bills while it works, not only when someone asks it a question. Teams that settle the hosting model first will find the model itself becomes interchangeable. Teams that adopt a vendor's agent first may find the hosting decision has been made for them.
The Enterprise Lens
Consider a distribution business where the customer service team spends the first two hours of every day chasing supplier confirmations and updating customers on delayed orders. An agent that does this overnight is now an off-the-shelf product rather than a custom build. The question is no longer whether AI can do the work. It is where that digital worker sits and what it is allowed to touch: your order system, your email, your supplier portals, your customer records.
A useful way to think about the three options is hiring. A vendor-hosted agent is like a temp from an agency: fastest to start, least control over how the work is done. An agent in your own cloud account is a contractor working inside your office under your rules. A self-hosted agent is a full-time employee you equip and supervise yourself. For most mid-sized firms, the sensible start is a vendor-hosted agent on a low-risk task that does not touch payments, contracts or sensitive customer data, with a clear plan for when the work justifies more control. The question to put to your technology partner is simple: if we move this agent somewhere else in a year, what do we lose?
What to Watch
- Whether OpenClaw Enterprise reaches 1.0 on schedule. The foundation has promised a reference architecture in the coming weeks. A credible production release signals that self-hosting agents is realistic beyond large IT departments.
- How Dots are priced beyond the first included agent. The cost of a second, tenth and fiftieth agent will determine whether vendor-hosted stays the cheapest route once usage grows.
- Whether other model providers ship "run it in your cloud" equivalents. If Google, Anthropic and Microsoft follow the Bedrock Managed Agents pattern, hosting location becomes a standard purchasing criterion, and the Private Inference preview due this autumn becomes the benchmark for data custody.
Sources
- Decrypt — OpenAI Gave AI Agents Their Own Computers at DevDay 2026
- Techzine — OpenClaw Enterprise brings AI agents safely to businesses
- Quartz — OpenClaw Foundation launches free, open-source enterprise control plane for AI agents
- Business Standard — OpenAI DevDay 2026: Dots agent, GPT-6.1 Sol, new plans