OpenAI launches Frontier: its bid to run the enterprise agent stack

OpenAI announced a new enterprise platform on Thursday called Frontier, built to let large companies deploy and manage fleets of AI agents across their internal systems. It is the company's clearest move yet from selling models to selling the layer those models run on — and the first customer list reads like a roll call of the Fortune 500: HP, Uber, Intuit, Oracle, State Farm, and Thermo Fisher are already on board.

The pitch is simple. Individual AI agents are easy to demo and hard to run at scale: they need access to company data, somewhere dependable to execute, a way to measure whether they did a good job, and guardrails that auditors and security teams can live with. Frontier bundles all four into one managed platform, aimed at companies that have finished experimenting with agents and now want to put them to work for real.

What Frontier actually is#

The platform is organized around four capabilities that enterprises currently stitch together themselves:

• Business context. A shared data layer that connects warehouses, CRM systems, ticketing tools, and internal applications, so agents reference the same information employees do — and build durable institutional memory over time instead of starting from zero every session.
• Agent execution. A dependable environment for reasoning over data, working with files, running code, and coordinating several agents in parallel — available on local infrastructure, enterprise clouds, or runtimes hosted by OpenAI.
• Evaluation and optimization. Built-in monitoring that tracks agent success rates, accuracy, and latency, plus feedback loops designed to sharpen behavior on real tasks rather than in a lab.
• Security and governance. Enterprise identity management for agent identities, auditable actions, and compliance with SOC 2 Type II and the ISO/IEC 27001, 27017, 27018, and 27701 standards, with companies choosing where their data sits at rest.

Frontier does not replace OpenAI's developer tools — the Agents SDK, AgentKit, and the API suite all stay. It sits above them as the management console: the place where a company sees what its agents are doing, directs them, and proves to auditors that they behaved.

The pitch: onboarding agents like people#

OpenAI is framing agent management on the way organizations scale their human workforce. In its telling, agents need the same things employees do: shared context, an onboarding process, hands-on learning with feedback, and clear permissions and boundaries. That HR metaphor is doing real strategic work — it tells CIOs this is a familiar management problem, not an alien one.

The ambition is sweeping. Fidji Simo, who runs OpenAI's applications business, has said that by the end of the year most digital work inside leading enterprises will be directed by people and executed by fleets of agents. OpenAI is also embedding deployment engineers inside early customers and working with a small cohort of partners — including Clay, Abridge, Harvey, Decagon, Ambience, and Sierra — to co-design industry-specific solutions on the platform.

Why now: the enterprise land grab#

Frontier arrives in the middle of a scramble for the enterprise agent stack. Anthropic released workflow-automation plugins for its Claude Cowork product last week; Microsoft is pushing its own agent-management play with Agent 365; AWS launched Bedrock AgentCore for companies that would rather assemble agent infrastructure from parts.

The awkward question is whether enterprises want a single vendor running the whole show. Tatyana Mamut, chief executive of the agent-observability startup Wayfound, told VentureBeat her clients are refusing to lock into multi-year commitments precisely because a better model could ship next month — they want the freedom to switch. OpenAI has not said whether Frontier will work with rival vendors' models and tools, and that silence is the biggest unanswered question of the launch.

Salesforce's AI chief Madhav Thattai, who oversees an agent-builder platform of his own, framed the stakes differently: the real value isn't in the model at all, but in the "last mile" — the software layer that converts raw capability into trusted, autonomous execution on a company's own data. Frontier is OpenAI's bid to own exactly that layer, and to own it across an entire enterprise rather than inside a single application.

The questions OpenAI did not answer#

Pricing is the headline unknown: OpenAI declined to disclose what Frontier costs, and gave no timeline for moving beyond its initial customer group to broader availability. BBVA, Cisco, and T-Mobile have piloted the approach, but the company has named only a handful of the dozens of businesses it says have tested the platform.

OpenAI's early evidence is its own. The company points to a financial-services customer that cut client-facing work time by 90%, and a technology customer saving 1,500 hours a month on product development — both OpenAI's chosen reference customers, both unverified by outsiders. Real proof will come only as Frontier deployments go public and independent numbers follow.

What to watch#

Three signals matter most from here. First, whether OpenAI opens Frontier to third-party models and tools — the answer decides whether this is a genuine platform or a well-designed moat. Second, what pricing looks like when it finally surfaces, because agent management sold as consumption could be cheap while the same capability sold as enterprise software could be enormous. Third, how Anthropic and Microsoft answer, since the enterprise market rarely tolerates a single winner for long. The era of the experimental AI pilot is ending. The fight now is over who runs the factory.