OpenAI acquires Ona (ex-Gitpod): Codex moves to persistent agents — tasks run even when the laptop is closed
🔎 Why OpenAI is buying the infrastructure for its own agents
On June 11, 2026, OpenAI announced the acquisition of Ona, a German startup rebranded after its spin-off from Gitpod. The news, revealed by CNBC, comes as Codex just passed the milestone of 5 million weekly users.
The problem this acquisition solves is simple: a coding agent that stops when you close your laptop is not an agent, it's an assistant. The true value of an autonomous agent lies in its ability to work continuously — running tests, refactoring a module, waiting for a code review — without the developer having to stay logged in.
OpenAI isn't just improving Codex. The company is moving up the stack: Layer 4 (models) is acquiring Layer 3 (execution infrastructure). This is a strategic move analyzed in detail by FourWeekMBA as a sign that the war of AI agents is now entering its infrastructural phase.
Anthropic led the way in May 2026 with its Claude Managed Agents, self-hosted sandboxes. OpenAI is responding with Ona, and this match changes the game for every developer using a coding agent today.
The essentials
- OpenAI acquires Ona (formerly Gitpod) on June 11, 2026, for an undisclosed amount, according to the official announcement.
- Codex reaches 5 million weekly users and gains the ability to run persistent tasks, even outside the developer connection.
- Ona becomes Codex's secure execution infrastructure layer: ephemeral, isolated cloud environments, with full lifecycle management.
- This is OpenAI's direct response to Anthropic's Claude Managed Agents (May 2026) and their self-hosted approach.
- InfoWorld analyzes this acquisition as a control move: OpenAI no longer wants to depend on third parties for executing its agents.
Recommended tools
| Codex | Persistent coding agent | Free / OpenAI Credits (June 2026, check on openai.com) | Developers wanting tasks that run 24/7 |
|---|---|---|---|
| Claude Code | Coding agent with self-hosted sandboxes | Anthropic Credits starting June 15, 2026 (check on anthropic.com) | Teams wanting infrastructure control |
| GitHub Copilot | Coding agent integrated into GitHub | Token billing (June 2026, check on github.com) | Developers already in the GitHub ecosystem |
What Ona actually is — and why it's not just a "cloud dev"
Ona is not just a simple cloud IDE. It's a platform for ephemeral, secure, and API-orchestratable development environments.
Its DNA comes from Gitpod: a workspace that is created in seconds, disappears after use, and isolates itself from everything else. Except that Ona specifically pivoted for AI agents. The difference is crucial: an environment for a human must display an interface. An environment for an agent only needs a terminal, a file system, and controlled network access.
InfoWorld points out that this acquisition allows OpenAI to "frame" its agents. Without controlled execution infrastructure, a coding agent can do anything — access sensitive files, open unauthorized network connections, consume resources without limit. Ona provides the sandboxing.
The approach echoes the manipulation learning challenges documented in research on dexterous systems (Learning Dexterous In-Hand Manipulation): an agent needs a controlled environment to learn and act effectively. Ona provides this environment to the world of code.
What persistence changes for a developer
Persistence is the transition from assistant to agent. Concretely, here is what becomes possible with the Ona integration in Codex.
Unsupervised long-running tasks
You give an instruction to Codex: "Refactor the authentication module, write the unit tests, run the CI, and fix failures in a loop until everything passes green." Without Ona, this task stops the moment you close your laptop or lose your connection. With Ona, the cloud environment continues executing the task.
The developer finds the result the next morning. It's a paradigm shift: we move from time spent with the agent to time saved by the agent.
Agent parallelism
A single developer can launch multiple simultaneous tasks on distinct Ona environments. One agent refactors the backend while another generates the API documentation, while a third fixes frontend bugs.
This logic of delegating to sub-agents is exactly the pattern described in Hermes Agent #14: Task delegation — orchestrating sub-agents. Ona provides the hardware infrastructure that makes this pattern operational at scale.
Reproducibility and isolation
Each Codex task runs in an isolated Ona environment. No dependency conflicts, no local system pollution, no "it works on my machine". The environment is reproducible: same versions, same config, same results.
For teams, this means that coding agents become predictable and auditable. A critical point for enterprise adoption.
The agent infrastructure war — OpenAI vs Anthropic
Until recently, the competition between OpenAI and Anthropic played out mainly on the models. GPT-5.5 dominates the agentic leaderboard with 98.2 points, followed by Gemini 3 Pro Deep Think at 95.4 and Claude Opus 4.7 (Adaptive) at 94.3. But models are no longer enough.
Claude Managed Agents: Anthropic's opening
In May 2026, Anthropic launches Claude Managed Agents — sandboxes where teams can host Claude Code's execution infrastructure themselves. The approach is "bring your own infrastructure": you control the cloud, network permissions, and environment lifespan.
This is appealing for regulated companies that don't want their codebases leaving their cloud perimeter. Claude Code becomes the go-to tool for this segment.
The Ona response: all-in-one at OpenAI
OpenAI takes the opposite path with Ona. No self-hosting, but a managed infrastructure natively integrated into Codex. You don't manage the workspaces — OpenAI does it for you.
FourWeekMBA clearly identifies the logic: by acquiring Ona, OpenAI internalizes Layer 3 to lock in the user experience around its Layer 4. The model and infrastructure become a cohesive package.
The Gartner MQ 2026 : OpenAI Codex, Cursor et GitHub Copilot leaders des agents de coding entreprise positions this integration as a key differentiating factor. Gartner notes that companies now evaluate coding agents not only on the quality of the generated code, but on the robustness of their execution infrastructure.
Impact on the coding agent market
The Ona acquisition is not just about OpenAI and Anthropic. It reshuffles the deck for all players in the coding agent space.
GitHub Copilot under pressure
GitHub Copilot remains the leader in terms of IDE integration, but its switch to token-based billing has created a wave of dissatisfaction among developers. The persistence offered by Ona+Codex adds a dimension that Copilot does not yet have: long-running execution offline.
Microsoft does have Azure Containers and GitHub Codespaces, but the integration is not as native as what Ona brings to Codex. The gap is widening when it comes to the pure developer experience.
Cursor in a tricky position
Cursor, the third player in the Gartner MQ 2026, relies on OpenAI models to function. With the acquisition of Ona, OpenAI strengthens Codex as a first-party product and mechanically reduces Cursor's competitive advantage. Why use Cursor with an OpenAI model if Codex offers better model + infrastructure integration?
Open source agents hit the infrastructure wall
The ecosystem of best autonomous AI agents and open source AI agents with Ollama locally is gaining maturity on the model side. LLMs like Kimi K2.6 (88.1) or GLM-5 (82) show that open source can compete. But the persistence infrastructure remains the weak point. Ona illustrates this gap: the best open source models still lack the cloud execution layer that Codex just gained.
The security question: why "framing" agents is essential
A coding agent running 24 hours unsupervised is also a security risk. This is precisely the point highlighted by InfoWorld in its analysis of the acquisition.
Sandbox by default
Ona applies the sandbox-by-default principle. Each Codex environment is isolated: no access to the host system, no leakage of secrets, no unauthorized network connection. The agent can only interact with the code assigned to it.
This is a fundamental shift from the early coding agents that ran directly in the developer's terminal, with all its access and secrets.
OpenAI o1's security model
The security approach is inspired by the mechanisms documented in the OpenAI o1 System Card, which details how OpenAI frames model reasoning to prevent undesirable behaviors. Ona transposes this principle to the infrastructure: what is framed at the model level is also framed at the execution level.
The limits of control
Despite sandboxing, a persistent agent remains a complex program. Research on the evaluation of high-level reasoning (A Systematic Assessment of OpenAI o1-Preview for Higher Order Thinking in Education) shows that even the most advanced models can produce flawed reasoning in certain scenarios. Applied to code, this means an agent can make poor decisions for hours without a human correcting course.
The trade-off is clear: more persistence requires more infrastructure guardrails. This is exactly what Ona brings.
The models behind Codex — what changes with Ona infrastructure
The integration of Ona does not modify the models themselves, but it radically changes the way their capabilities are leveraged.
GPT-5.3 Codex in context
GPT-5.3 Codex, with its score of 80 on the agentic leaderboard, is not OpenAI's most powerful model. GPT-5.5 (98.2) and GPT-5.4 Pro (91.8) outperform it. But Codex is specifically optimized for code generation and repair.
Research on automatic program repair with Codex (Automatic Program Repair with OpenAI's Codex: Evaluating QuixBugs) already demonstrated bug-fixing capabilities back in 2021. With Ona, these capabilities can operate in a closed loop: the agent fixes, runs tests, identifies failures, fixes again — without human intervention.
Scaling with reasoning models
For complex tasks requiring in-depth reasoning, Codex can now rely on o1-preview (90.2) or GPT-5.4 (87.6) in the background. The Ona environment allows scaling up on the model used without fearing that a long task will be interrupted.
The study on using Codex for HPC kernel generation (Evaluation of OpenAI Codex for HPC Parallel Programming Models Kernel Generation) showing the limits of Codex on complex parallelized code takes on a new meaning here: with a persistent environment, the agent can iterate much longer on these difficult problems.
Claude Opus 4.7 vs GPT-5.5 in persistent execution
The battle of the best LLMs for AI agents takes on a new dimension. Claude Opus 4.7 (94.3) in a self-hosted Claude Managed Agent vs GPT-5.5 (98.2) in a managed Ona environment — this is no longer just a model comparison, it's a comparison of complete stacks.
Computer Use and the GUI/API frontier — what Ona's infrastructure makes possible
An often underestimated aspect of the Ona acquisition is its connection to the evolution of Computer Use. Coding agents are no longer limited to the terminal: they interact with graphical interfaces, browsers, APIs.
From the headless environment to the full GUI
Ona can provision environments with or without a graphical interface. For a pure coding agent, a terminal is enough. But for an agent that needs to test a web application in a real browser, navigate a staging interface, or validate a visual render — an environment with a GUI is necessary.
The research ToolCUA : quand les agents Computer Use apprennent à choisir entre GUI et API explores exactly this frontier: an agent capable of dynamically deciding whether it interacts via API (fast, reliable) or via GUI (necessary for certain validations). Ona provides the infrastructure that makes both modes possible within the same persistent workspace.
Implications for testing workflows
A Codex agent in an Ona environment can: compile the code via the terminal (API/CLI), launch the application, open it in a headless browser, verify the visual render, and document everything — all without any human being connected. This is the end-to-end agent the industry has been calling for for months.
What this acquisition means for the AI stack in general
Beyond coding, OpenAI's acquisition of Ona signals a broader movement in the AI ecosystem.
Vertical consolidation is accelerating
Layer 4 buys Layer 3. AI model companies no longer want to be mere API providers. They want to control the entire chain: model, orchestration, execution, monitoring.
This is the same logic that drove Apple to integrate its chips, its OS, and its services. OpenAI is reproducing this model in the AI cloud.
Pure-play agent infrastructure players are threatened
Startups positioning themselves as "the infrastructure for AI agents" — workspace provisioning, lifecycle management, sandboxing — find themselves facing a giant that is internalizing this layer. If OpenAI successfully integrates Ona, other acquisitions will likely follow from Google (for Gemini) or xAI (for Grok 4.1 and beyond).
Open source must respond
The open-source ecosystem, with models like Kimi K2.6 in self-host or GLM-5 in reasoning mode, offers a credible alternative at the model level. But without an Ona equivalent, the developer experience falls short. Projects like Devcontainers combined with open-source agent orchestrators could bridge this gap, but the road is long.
❌ Common mistakes
Mistake 1: Confusing Ona with a simple cloud IDE
Ona is not Gitpod with a new name. It is an execution platform for agents, without a mandatory human interface, designed to be orchestrated via API. Comparing it to GitHub Codespaces is like comparing a datacenter to a coworking space.
Mistake 2: Thinking that persistence solves all problems
An agent running 24h unsupervised can also accumulate errors for 24h. Persistence amplifies both successes and failures. Monitoring and safeguards (which Ona integrates) are just as important as persistence itself.
Mistake 3: Believing Claude Code is dead
Claude Code is still very much alive, particularly in companies that require self-hosting for compliance reasons. Anthropic's approach (you control the infrastructure) and OpenAI's approach (Ona controls the infrastructure for you) target different segments. This is not a zero-sum game.
Mistake 4: Ignoring the cost of persistence
Running a cloud environment for 12 hours costs money. Persistence is not free, and models like GPT-5.5 consume tokens at every iteration. The real challenge for OpenAI will be making this persistence economically viable, including for individual developers.
❓ Frequently Asked Questions
What exactly is Ona?
Ona is a German startup, spin-off of Gitpod, that provides ephemeral and secure cloud environments designed specifically for running AI agents, with no human interface required.
Do Codex tasks really run offline?
Yes. With the Ona integration, the execution environments are hosted in OpenAI's cloud. The developer can disconnect, close their laptop, and retrieve the results later.
How is this different from Anthropic's Claude Managed Agents?
Claude Managed Agents (May 2026) is self-hosted: you provide the infrastructure. Ona is managed by OpenAI: it's an integrated service. Two opposing philosophies to solve the same problem.
Does this acquisition also concern non-coding agents?
For now, the Ona integration is focused on Codex. But the platform is generic — it could serve as a foundation for other types of OpenAI agents in the future.
Which Codex model does it use with Ona?
GPT-5.3 Codex (agentic score: 80) is the default model, but the architecture allows for using other OpenAI models like o1-preview (90.2) or GPT-5.4 (87.6) depending on the complexity of the task.
✅ Conclusion
OpenAI's acquisition of Ona marks the moment when the war of coding agents moves from the realm of models to that of infrastructure. An agent that stops when you close your laptop is just an assistant; an agent that continues without you is a true worker. Codex + Ona makes this transition. Anthropic responds with self-hosting, GitHub Copilot with token billing, and open-source agents are looking for their equivalent of Ona. The next battle will not be won on the benchmark, but in the datacenter. To follow the evolution of the best autonomous AI agents in this new infrastructure landscape, model rankings are no longer enough — you have to look at the complete stack.