Why OpenAI Bought Ona: Cloud Execution for AI Agents

OpenAI is acquiring Ona — the German company formerly known as Gitpod — to give Codex the one thing a long-running AI coding agent cannot live without: a persistent cloud environment that keeps running for hours or days, even after the developer closes their laptop. The deal was announced on June 11, 2026, with financial terms undisclosed.

Most coverage led with the logo swap. The more interesting story is the architectural constraint that forced the purchase. Today’s coding agents are session-bound: spin one up in your editor, and the work stops the moment the session ends. As agents move from minutes-long edits to multi-hour and multi-day tasks, that session-based model breaks — and the industry has no agreed answer for where that work should actually execute.

This analysis covers what OpenAI bought, the durable-execution problem that explains the timing, who Ona is, why Codex’s growth made the deal urgent, and the build-vs-buy decision every team building agents now faces between neutral sandboxes — E2B, Modal, Daytona — and a first-party Codex Cloud runtime. Every figure below is sourced and attributed.

On June 11, 2026, OpenAI announced it will acquire Ona, the cloud development platform formerly known as Gitpod, to support longer-running Codex agent tasks. Per CNBC’s reporting, financial terms were not disclosed, and the deal is subject to customary closing conditions, including regulatory approvals. Ona CEO and co-founder Johannes Landgraf, along with the full team, will join OpenAI’s Codex group once the deal closes.

What OpenAI is buying is not a model or a dataset — it is an execution layer. Ona’s technology provides three core components: sandboxed cloud developer environments, AI-powered agents with workflow automation, and enterprise-grade security guardrails (role-based access control, single sign-on, OIDC, audit logging, and deployment inside a customer’s own VPC). That last piece is the part you cannot easily build in a quarter — the compliance and isolation plumbing that lets regulated industries run autonomous agents at all.

Here is the architectural fact that explains the entire deal. Most of today’s AI coding agents run inside an editor session or a short-lived process. Start a task, and the agent works against your local machine or a transient container. Close the laptop, lose the network, or let the session time out, and the work stops. For a two-minute refactor that is fine. For a task that needs to run for three hours — migrate a service, work through a backlog of failing tests, refactor across a large codebase — it is fatal.

This is the “single-core to multicore” moment for agentic coding. The shift is from session-based execution, where an agent lives and dies with your editor, to durable cloud execution, where an agent gets its own persistent environment that keeps running independently. Ona’s 10-minute automatic checkpointing is a direct answer: the agent’s progress is saved continuously, so a long task survives a dropped connection or a closed lid. For deeper background on how agents get isolated workspaces, see our guide to agent sandboxing and isolation patterns.

Landgraf made that point at the September 2025 rebrand, well before the OpenAI deal — which is exactly why it matters. The pivot from a human-facing IDE company to an agent-execution company was a bet that the bottleneck for autonomous coding would move from intelligence to infrastructure. The acquisition is the market agreeing with that bet. As Landgraf put it in the context of the deal, agents need more than intelligence; they need a trusted workplace — a durable, secure, isolated environment to actually do the work.

Ona is headquartered in Kiel, Germany, and was founded around 2019–2020 as Gitpod — an open-source, browser-based cloud development environment that let engineers spin up a ready-to-code workspace from any Git repository. In November 2022 it raised a $25 million Series A led by GitHub co-founder Tom Preston-Werner, with General Catalyst, Crane Venture Partners, Speedinvest, and Vertex Ventures US joining, alongside angel investments from Shopify CEO Tobi Lütke and Datadog CEO Olivier Pomel. By its September 2025 rebrand, Gitpod reported serving 2 million developers globally (a vendor-stated figure).

On September 2, 2025, Gitpod rebranded to Ona and repositioned entirely — from cloud dev environments for humans to AI agent orchestration infrastructure. The legacy path was wound down in parallel: Gitpod Classic’s pay-as-you-go service shut down on October 15, 2025, with the self-hosted Gitpod Flex replacing it for non-AI users. The new Ona stack centers on giving agents the same sandboxed, reproducible environments Gitpod once gave developers.

The timing is not coincidental. Per OpenAI, Codex now supports more than 5 million weekly active users as of June 2026, up from 3 million in April 2026 — and OpenAI frames that as a more-than-400% increase since early 2026 and more than a 6× jump since the Codex desktop app launched in February 2026. (The 400% and 6× figures are measured from different baselines — early 2026 versus the February desktop launch — so they describe two different windows, not a single rate.) When a product’s usage is compounding that fast, the execution infrastructure underneath it becomes the constraint, and buying a proven one beats building it under load.

The more strategically interesting shift is who is using Codex. Knowledge workers — non-developers — now make up about 20% of Codex users, per OpenAI, and are growing at roughly 3× the developer adoption rate. The fastest-growing non-developer tasks are data analysis, research, and the creation of knowledge artifacts like reports and spreadsheets. (OpenAI cites week-over-week growth figures for these categories in its own “knowledge work” report; those are vendor-stated and worth treating as directional.) For the full growth picture, our earlier breakdown of Codex’s growth trajectory tracks the same curve from a developer lens.

That non-developer surge reframes the Ona deal. If a fifth of Codex usage is people who do not manage their own development environment — analysts, researchers, operators — then “the agent must run somewhere reliable that I never have to configure” stops being a convenience and becomes a requirement. A managed, durable cloud runtime is what makes Codex usable for someone who has never opened a terminal. Ona supplies exactly that layer, with enterprise security already wired in. OpenAI has been assembling this stack deliberately — it also acquired Astral, makers of the Python tools uv and Ruff, earlier in 2026 to expand Codex’s Python ecosystem.

Once you accept that long-running agents need a durable place to run, the next question is where. There is now a real market of execution runtimes, and they split into two camps: neutral, model-agnostic sandboxes you can point at any model, and first-party runtimes tied to a single vendor’s agent. The table below is ours — built from published session limits, isolation technology, and an editorial read of lock-in risk. No single published comparison combines all three, and that combination is what a dev lead actually needs when choosing a cloud execution layer in mid-2026.

Two patterns jump out. First, session caps matter more than they look for agent work. Vercel Sandbox, which reached general availability in January 2026, caps runtime at 45 minutes on Hobby and 5 hours on Pro/Enterprise — fine for short tasks, but a hard wall for a multi-hour migration. E2B raises that ceiling to 24 hours per session with pause-and-resume, while Modal, Daytona, and Northflank advertise unlimited session duration, which is what genuinely day-long agent workloads require. Cursor’s cloud agents, launched February 12, 2026, run in isolated Linux VMs reported to sustain 25 to 52-plus hours of continuous autonomous work. For more on how the underlying functions and sandboxes are built, see our guide to serverless and sandbox execution environments and the deeper look at Cursor’s cloud agent implementation.

Here is the practitioner angle the announcement coverage skips. With Cursor launching its own cloud VMs and OpenAI acquiring Ona, the two largest agent vendors now own their execution layers. That is convenient if you have standardized on one of them. It is a strategic risk if you have not. A first-party runtime quietly couples two decisions that used to be separate: which model you run, and where it runs. Once your agents live inside Codex Cloud or Cursor’s VMs, switching models means switching execution platforms too.

That is precisely why the neutral providers — E2B, Modal, Daytona, Northflank — still have a clear role. They let you keep the execution layer model-agnostic, so you can route a task to Codex today and a different model tomorrow without re-platforming. The trade-off is real in both directions: first-party runtimes give you a tightly integrated, lower-configuration experience; neutral ones give you portability and negotiating leverage. Teams running genuine multi-vendor strategies should weight portability heavily, because the cost of lock-in compounds as agent usage scales. This is the same logic we apply when designing multi-agent orchestration patterns and when wiring Codex’s multi-agent capabilities into a production stack.

One detail most coverage mentions but does not analyze: Ona is a German company, and the deal is explicitly subject to regulatory approvals as part of its closing conditions. A German headquarters makes a review by the Bundeskartellamt — Germany’s federal antitrust authority — the likely path. No timeline or outcome has been published, so the honest framing is that closing is conditional, not certain, and that the regulatory step adds a genuine (if probably modest) risk to the deal calendar. We are not predicting an outcome; we are noting that “announced” and “closed” are different states, and the gap between them runs through a regulator.

For teams making roadmap decisions, the practical implication is simple: do not assume Codex Cloud durable execution is a shipping, generally-available feature the day after the announcement. An acquisition that must clear regulatory review can take months to close, and the product integration follows the close, not the press release. Plan your own execution-layer choices on what is available today, and treat the Ona-powered Codex runtime as a roadmap signal rather than a current capability.

The deal changes the decision tree for any team building or operating long-running agents. Below is how we would reason about the execution layer right now, by team profile.

Our broader read: the execution layer is becoming as strategic as the model choice itself, and most teams are not treating it that way yet. For the next year, the pragmatic move is to keep the runtime decision reversible — favor a neutral sandbox for anything multi-vendor or multi-hour, and adopt a first-party runtime only where the integration genuinely outweighs the lock-in. If you are weighing these trade-offs for your own agent stack, our AI digital transformation engagements start with exactly this kind of build-vs-buy evaluation, and our development team can stand up the runtime and orchestration to match.