OpenAI has agreed to buy Ona, a Kiel, Germany-based cloud infrastructure company, so its Codex coding assistant can run agents that keep working for hours or days at a time. The deal was announced June 11, 2026, and it’s still waiting on regulatory sign-off. If you use Codex, or you’re trying to figure out whether to build your own agent stack instead of renting one, here’s what this actually means for you.
Quick Verdict
Nothing changes for you today. The deal hasn’t closed yet, and OpenAI said it’s still subject to standard regulatory approval. What it signals matters more than what it does right now: OpenAI just admitted, in public, that the biggest problem with AI agents isn’t model intelligence anymore it’s infrastructure. You prompt something, you close your laptop, and the agent stops. Ona fixes that specific problem. If you’re already deep into Codex, expect persistent cloud environments to show up as a feature sometime after closing, not immediately. If you’re building agents yourself with something like Agent Zero, this is a signal worth paying attention to, not a reason to switch tools today.
What Ona Actually Builds
Ona gives AI agents secure, pre-configured cloud environments stocked with the tools, access controls, and context needed to keep working over time. It started life in 2019 as Gitpod, a company known for spinning up cloud-based dev environments so engineers didn’t have to configure their laptops from scratch every time they joined a new project. Somewhere along the way, the founders realized the same problem “I need a consistent, secure place to do work that isn’t tied to one machine” applies just as much to AI agents as it does to human developers.
The company rebranded to Ona and pivoted toward serving AI agents specifically. Its enterprise client list reportedly includes a major U.S. bank, several European pharmaceutical companies, and sovereign wealth funds in Asia. According to MLQ News, productive use of Ona’s agent environments among enterprise clients grew 13-fold in 2026 alone. That’s not a vanity metric it tells you real companies were already paying for this before OpenAI showed up with a checkbook.
Why OpenAI Bought It (Before the How)
Here’s the thing most coverage skipped: this acquisition isn’t really about Codex getting smarter. It’s about Codex getting less forgetful.
Right now, most AI coding tools Codex included work in session-based bursts. You send a prompt, the agent responds, the context window fills up, and eventually you start over. That’s fine for quick fixes. It falls apart completely for the work that actually ships software: scanning an entire codebase for vulnerabilities, running a multi-day application modernization project, or managing a data migration that takes longer than your laptop battery.
OpenAI’s own announcement put it plainly: Codex’s most valuable work now unfolds over hours or days, not minutes, and people shouldn’t have to stay tied to the machine where a task began. That’s a real shift in how the company is thinking about its product. Early Codex was a tool for developers who wanted help writing code while they sat at their desk. The Codex OpenAI is building toward now is something closer to a digital employee you can hand a multi-day project to and check back on later.
I’ve tested a fair number of “autonomous” coding agents over the past year, and the failure pattern is almost always the same: the agent does great work for 20 minutes, then loses context, forgets what it already tried, or just stops because the session timed out. That’s not a model problem. That’s an infrastructure problem. Ona’s pitch persistent, customer-controlled cloud environments is a direct answer to exactly that failure mode. Whether it works as smoothly in production as it does in the announcement is the part nobody can verify yet.
There’s also a more obvious business reason. Codex usage numbers have moved fast: more than 5 million weekly users as of June 2026, up from roughly 3 million in April, which OpenAI describes as a 400% increase since earlier in the year. Knowledge workers outside of software engineering now make up around one in five Codex users, and that segment is growing about three times faster than the core developer base. OpenAI didn’t build Codex to stay a coding tool. It’s building it to be the agent layer for entire companies, and that requires infrastructure that traditional AI labs don’t usually own: secure cloud execution, audit trails, access controls, the boring enterprise plumbing that actually gets deals signed.
How the Deal Actually Works
The mechanics are fairly standard for this kind of acquisition, with one detail worth noting. OpenAI is buying Ona’s execution and orchestration layer, not replacing it with something proprietary from scratch. Once the deal closes, Ona’s team — led by CEO Johannes Landgraf joins OpenAI’s Codex division directly, and Ona’s customer-controlled execution model gets folded into Codex’s intelligence and orchestration layer.
“Customer-controlled” is the operative phrase here. The idea is that agents run inside an organization’s own cloud environment, with OpenAI providing the model and the coordination logic, rather than everything running inside OpenAI’s infrastructure by default. For enterprise buyers worried about data residency or compliance — banks, pharma companies, government-adjacent clients that distinction is the entire reason a deal like this gets approved internally. Nobody signs off on sensitive workloads running somewhere they don’t control.
Financial terms weren’t disclosed, which is typical for deals at this stage, and it’s still pending regulatory approval. Until it closes, OpenAI and Ona remain legally separate companies. That’s worth remembering before anyone tells you Codex now has Ona’s tech built in — it doesn’t, not yet.
The Numbers Worth Knowing
A few figures from this announcement are going to get cited everywhere, so it’s worth having the actual context instead of the headline version:
Codex crossed 5 million weekly active users in June 2026, up from about 3 million in April — roughly a sixfold jump since OpenAI launched its desktop app in February. Knowledge workers (sales, investment banking, equity research, and similar roles) now represent close to 20% of Codex’s user base and are growing at triple the rate of the original developer audience. On the Ona side, enterprise agent usage grew 13-fold during 2026, with deployments inside at least one major U.S. bank and multiple pharmaceutical and sovereign-wealth clients.
Put those together and you get the actual strategic story: OpenAI isn’t buying Ona because Codex needs more users. It’s buying Ona because the type of user is shifting toward people who need agents to run securely, for long stretches, inside corporate infrastructure that has real compliance requirements. That’s a different product than “help me write a function.”
This Isn’t OpenAI’s First Acquisition Rodeo
If this is the first OpenAI acquisition you’ve noticed, it’s worth knowing it’s part of a pattern, not a one-off. In May 2025, OpenAI bought Jony Ive’s hardware startup io for more than $6 billion, signaling ambitions well beyond software. In October 2025, it acquired Software Applications, the team behind an AI interface called Sky built for Apple Mac users. In January 2026, it picked up healthcare tech startup Torch for roughly $60 million. In March, it announced the purchase of Promptfoo, a cybersecurity startup focused on testing AI systems for vulnerabilities. Ona is the latest in that line, and arguably the most directly tied to Codex’s core product roadmap.
There’s also a parallel move worth mentioning: OpenAI recently partnered with Visa to let AI agents handle financial transactions directly, integrating Visa’s payments network and tokenization technology so agents can make purchases on a user’s behalf with proper security controls. Read those two announcements together and a clearer picture forms — OpenAI is trying to build agents that can both do persistent work (Ona) and transact in the real world while doing it (Visa). That’s a meaningfully bigger ambition than “better autocomplete for code.”
The Anthropic Angle Nobody Should Skip
You can’t read this acquisition in isolation from the competitive dynamics. Anthropic’s Claude Code has been growing fast and is widely credited as a major driver of Anthropic’s broader momentum over the past year. Both OpenAI and Anthropic have reportedly filed confidential prospectuses with the U.S. Securities and Exchange Commission as they consider public offerings, possibly as early as this fall.
This matters for a simple reason: when two companies racing toward a possible IPO both lean hard into “AI coding agents” as a flagship product category, infrastructure acquisitions like Ona stop looking optional. Owning the execution layer — rather than renting it from a generic cloud provider becomes a competitive moat. OpenAI’s bet, based on its own public statements, is that the limiting factor for agents isn’t model capability anymore. It’s whether an agent can keep working reliably, securely, and continuously without a human babysitting the session. Whoever solves that cleanly first gets the enterprise contracts. That’s the real race happening underneath the headline.
What This Means If You’re Already Using Codex
Honest answer: not much changes this week. The acquisition hasn’t closed, and OpenAI hasn’t published a timeline for when Ona’s infrastructure actually shows up inside Codex for regular users. A few practical things worth doing in the meantime:
If you’re running long Codex sessions today and hitting context loss or session timeouts, that’s the exact pain point Ona is meant to solve eventually but don’t restructure your workflow around a feature that doesn’t exist yet. Keep using whatever session-management tricks you’ve already got (checkpointing your work, breaking big tasks into smaller scoped requests) until there’s an actual product update to react to.
If you’re an enterprise buyer evaluating Codex against alternatives, ask your OpenAI rep directly about timeline and what “customer-controlled cloud environment” will actually mean for your compliance team. Vague promises about future infrastructure shouldn’t factor into a procurement decision today.
If you’re a developer who’s curious about persistent agent environments right now, rather than after this deal closes, it’s worth looking at what’s already available. Our guide to setting up Agent Zero covers a self-hosted approach that gives you a comparable persistent-environment setup without waiting on OpenAI’s roadmap.
What This Means If You’re Building Your Own Agent Stack
This is where I’d push back a little on the breathless “this changes everything” framing you’ll see elsewhere. If you’re already running your own agent infrastructure say, with Agent Zero, AutoGPT-style frameworks, or a custom orchestration layer — this acquisition doesn’t make your setup obsolete. It validates the architecture you already chose.
The core insight behind Ona (and Gitpod before it) isn’t proprietary or new: agents need a persistent, isolated, properly permissioned environment to do real work, not just a chat session with tool-calling bolted on. That’s exactly the problem self-hosted agent frameworks have been solving the hard way for over a year. If you’ve gone through the process of setting up Agent Zero inside Docker, you’ve basically built a smaller, DIY version of what Ona sells to banks and pharma companies minus the enterprise sales team and the SOC 2 paperwork.
The honest tradeoff: a managed solution like what Codex-plus-Ona will eventually offer saves you the operational headache of managing your own cloud environments, security boundaries, and uptime. The cost is control and, usually, price. If you’re a solo developer or small team, self-hosting still wins on flexibility and cost. If you’re an enterprise with compliance requirements and no internal infra team to dedicate to this, a managed offering starts looking a lot more attractive which is precisely the buyer OpenAI is targeting with this purchase.
For people weighing both paths, it’s worth comparing setups side by side. Our step-by-step walkthrough for installing Agent Zero with Docker and our broader Agent Zero overview guide both cover the self-hosted route in detail. If you’d rather see the bigger picture of building agents from scratch instead of relying on a vendor’s roadmap, our tutorial on building autonomous AI agents walks through the architecture decisions you’ll actually face. There’s also a newer option worth knowing about if Agent Zero isn’t quite the right fit our setup guide for the Neo AI agent framework covers an alternative approach with different tradeoffs around persistence and orchestration.
Security Implications Worth Taking Seriously
Here’s the part that doesn’t get enough attention in most coverage of this deal: persistent agents that run inside your cloud environment for hours or days are a meaningfully bigger attack surface than a chatbot that responds and forgets. An agent with standing access to your codebase, your credentials, and your infrastructure for an extended period is a different security model than a session that ends when you close the tab.
OpenAI’s framing — “customer-controlled execution” with audit trails and access boundaries is the right instinct. Whether it holds up under real adversarial testing is a separate question entirely, and one that security teams should be asking before they greenlight any persistent-agent deployment, regardless of vendor. This is exactly the kind of scenario driving demand for people who specialize in stress-testing AI systems before they go into production. If that’s a career direction you’re curious about, our breakdown of AI red team roles and what the job actually involves is worth a look, especially given how fast this specific niche is growing right now.
What This Means for the Broader Job Market
Acquisitions like this one are a decent leading indicator of where hiring demand is headed. When a company the size of OpenAI spends real money buying persistent-agent infrastructure, it usually means they’re staffing up around deploying agents in production, not just building flashier demos. That tends to create downstream demand for roles focused on agent orchestration, deployment, and monitoring distinct from the pure model-research jobs that dominated headlines a couple years ago.
If you’re tracking where the actual openings are showing up, our rundown of current agentic AI job roles covers the specific titles and skill sets companies are hiring for right now, which is a more useful signal than guessing from press releases.
A Quick Note on Where Agent Infrastructure Is Heading Generally
It’s worth zooming out for a second. The pattern in this acquisition generalized infrastructure originally built for one use case (developer cloud environments) getting repurposed for AI agents isn’t unique to coding. The same persistent-environment, secure-execution logic applies anywhere an agent needs to operate continuously and access sensitive systems, whether that’s software development, financial analysis, or more specialized verticals. Even somewhat unexpected domains, like the data-heavy modeling work behind AI-driven sports betting models, run into the exact same infrastructure problem: agents that need to crunch data continuously and make decisions over extended periods, not just answer a single prompt and stop. The Ona acquisition is really a bet that this underlying infrastructure problem is universal across verticals, not specific to coding.
Prompting Agents That Actually Run for Hours
One thing that changes once agents can run persistently: how you prompt them matters a lot more. A quick one-shot prompt works fine for a task that finishes in two minutes. It falls apart for something running across six hours, because the agent needs enough context, constraints, and checkpoints built into the original instruction to avoid drifting off-task by hour four.
If you’re going to be working with longer-running agents whether that’s a future version of Codex with Ona’s infrastructure, or a self-hosted setup today it’s worth getting deliberate about prompt structure now rather than after something goes wrong on a multi-hour run. Our guide to advanced prompt engineering techniques covers the structural approaches that hold up over longer agent sessions, and our collection of Agent Zero prompt examples shows specifically how that translates into real prompts for persistent agent tasks.
The Honest Risks Nobody’s Highlighting
A few things worth flagging that the celebratory coverage mostly skipped:
This deal hasn’t closed. Regulatory approval isn’t guaranteed on any timeline, and “subject to customary closing conditions” is corporate language for “this could take a while or fall apart.” Don’t build a six-month roadmap around a feature that depends on a deal that hasn’t finished yet.
Integration risk is real. I’ve watched enough acquisitions limp through a rocky 12-to-18-month integration period to be skeptical of “seamless” framing in announcement blog posts. Ona’s tech was built for a different primary use case (developer environments) before pivoting to agents. Bolting that into Codex’s existing architecture is a real engineering project, not a flip of a switch.
There’s also a quieter risk worth naming: as OpenAI consolidates more of the agent stack — model, orchestration, and now execution environment under one company, customers lose some of the leverage that comes from mixing vendors. That’s a fine tradeoff if you trust the platform completely. It’s worth a second thought if you don’t.
What to Actually Do With This Information
If you’re a Codex user, there’s nothing to set up yet just keep an eye out for an announcement once the deal closes and Ona’s infrastructure actually ships. If you’re evaluating whether to build versus buy your agent infrastructure, treat this acquisition as a strong signal that persistent execution environments are the next real battleground, and decide now whether you want to own that layer yourself or rent it from whichever AI lab wins the race. Either path is defensible. What’s not defensible is ignoring the shift entirely and assuming session-based agents are good enough for the work you’ll be asking them to do a year from now.
For more on how agent infrastructure and tooling are evolving heading into the second half of 2026, our full coverage archive is at theaijournal.co.
