What you need to know
- The deal: announced on 11 June 2026, OpenAI is acquiring Ona — the German cloud startup formerly known as Gitpod (Gitpod GmbH). Terms were not disclosed.
- What Ona does: it runs Codex agents inside customer-controlled cloud sandboxes that can operate unattended for hours or days at a time.
- The team: Ona's people, including chief executive Johannes Landgraf, join OpenAI's Codex group.
- The scale: Codex now has more than five million weekly users, up roughly 400% earlier this year.
- The shift: coding agents move from short interactive sessions to long-running, sandboxed, autonomous execution — and "agent ops" becomes a skill worth hiring for.
If you only learn one thing from this acquisition, make it this: the value is no longer in the chat box, it is in the sandbox. Get comfortable scoping a task, handing it to an agent that runs for a long time on its own, and then reviewing the diff it produces. That review-and-verify loop is the work that will still be yours in 2027.
What was announced
On 11 June 2026, OpenAI confirmed it had agreed to acquire Ona, the cloud development company that most builders still know by its previous name, Gitpod. The deal terms were not disclosed. OpenAI framed the move on its own blog as a way to scale Codex — its coding-agent product — towards genuinely long-running, autonomous work, and the announcement was picked up the same day by CNBC, The Next Web and How2Shout among others.
Gitpod spent years building cloud development environments: reproducible, ephemeral workspaces that spin up on demand so a developer never has to fuss with a laptop's local toolchain. Under the Ona brand, the company pivoted that same machinery towards agents. Instead of a human typing into a hosted workspace, Ona runs Codex agents inside isolated cloud sandboxes — environments the customer controls — and lets them grind away unattended. The headline capability is duration: these agents can keep working for hours, and in some cases days, without a person babysitting the session.
That is the part worth sitting with. Most builders' mental model of a coding agent is still a conversation: you ask, it responds, you nudge, it responds again. Ona's pitch — and now OpenAI's — is that the conversation is the bottleneck. A long task should be handed off, not chatted through.
"Runs for days" is a capability, not a promise of correctness. An agent that works unattended for forty-eight hours can also drift, loop, or quietly rack up cloud and token costs for forty-eight hours. The sandbox isolation matters precisely because the failure modes are now longer and more expensive than a bad reply in a chat window.
Why OpenAI wanted it
The strategic logic is straightforward once you look at the numbers. OpenAI says Codex now has more than five million weekly users, up roughly 400% earlier this year. That is a product with real momentum, and the natural next question for any agent product is: how long and how independently can it run? Short interactive sessions are easy to host. Long-running, isolated, stateful sandboxes that customers trust with their source code are hard — and that is exactly the muscle Gitpod spent years building.
Buying Ona gives OpenAI three things at once. First, the orchestration layer: the plumbing that launches, isolates, snapshots and tears down sandboxes safely. Second, the team that knows how to operate it at scale, including chief executive Johannes Landgraf, who moves into the Codex group. Third, a credibility story for enterprises that will not let an agent touch their repositories unless the execution happens in an environment they control. For a dual-market reader, that last point lands hard: a fintech in Mumbai or a regulated insurer in Manchester both need "the agent ran in our cloud, on our terms" before they will sign anything.
From chat sessions to long-running agents
It helps to see the two models side by side, because the difference in working style is larger than it first appears.
| Dimension | Interactive coding assistant | Long-running sandboxed agent |
|---|---|---|
| Session length | Seconds to minutes | Hours to days |
| Human role | Prompt, nudge, accept each step | Scope upfront, review the result |
| Execution venue | Editor or chat window | Customer-controlled cloud sandbox |
| Failure surface | A bad suggestion you can ignore | Hours of drift, loops, or cost |
| Cost model | Per-turn tokens | Tokens plus compute, over time |
| Skill that matters | Prompting | Scoping, guardrails, verification |
The left column is the world most teams optimised for in 2024 and 2025. The right column is where the Ona deal points. None of this makes the interactive assistant obsolete — short turns are still better for exploratory work — but it does mean the centre of gravity moves towards orchestration. We covered the broader contest between these products in our look at Claude vs OpenAI Codex vs Gemini for computer-use agents, and the stack-level choices in choosing your 2026 coding-agent stack.
"The mental shift took me a fortnight. I stopped asking 'what do I type next' and started asking 'what would I check if a junior engineer disappeared for two days and came back with a branch'. That question is the whole job now — write the spec well, set the boundaries, then review like a maintainer."
— A Verified Builder · Bengaluru, INWhat it means for India and UK builders
Here is the practical part. If your day is largely typing code that an agent could plausibly produce, the long-running model is both a threat and an opportunity — and which one it becomes depends on what you learn next. The premium is migrating from authorship to supervision. Specifically:
- Task scoping becomes a first-class skill. An agent that runs for a day needs a brief that is unambiguous about scope, acceptance criteria and what it must not touch. Vague prompts that work fine in a chat window produce a day of expensive wandering in a sandbox.
- Guardrails and isolation are now part of the engineering job, not an ops afterthought. Knowing how to box an agent into a sandbox with the right permissions, secrets handling and network limits is increasingly what separates a hobby setup from a production one.
- Diff review at volume is the new code review. When an agent returns a 4,000-line branch after running overnight, the skill is reading it critically, fast, and catching the subtle wrong thing — not re-deriving it.
- Cost and run-time literacy matters in a way it did not for per-turn chat. A long run touches both tokens and compute over time; builders who can reason about that economics will be trusted with the budget.
For builders in India and the UK specifically, this is levelling in a useful way. A long-running agent does not care whether its supervisor sits in Pune or Bristol. The work that is hardest to offshore — clear specification, careful review, accountability for what shipped — is exactly the work this shift rewards. The risk is the inverse: roles defined purely as "produce the boilerplate" compress fastest. The move is to climb the ladder towards agent ops before the rung you are standing on gives way. Our guide on getting more from AI coding agents is a sensible starting point, and the broader SDK landscape is mapped in our agent-SDK wars piece.
Spend a weekend running one real, scoped task in a sandboxed agent end to end — brief, run, review, ship. Then write up what broke and what you changed. That single artefact tells a hiring manager more about your readiness for the long-running-agent era than any certificate, and it gives you something concrete to put on a profile.
Every article here is written for working Builders. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The infrastructure angle
Strip away the product framing and this is, at heart, an infrastructure acquisition. Long-running agents need somewhere safe to run, and "somewhere safe" is not a trivial problem. You need reproducible environments, fast cold starts, snapshots so a run can be paused and resumed, hard isolation between tenants, secrets that never leak into a model's context, and observability so you can see what a day-long run actually did. Gitpod built most of that for human developers; Ona repointed it at agents; OpenAI now owns it.
That is also why "customer-controlled" is the phrase doing the heavy lifting. Enterprises in regulated sectors — finance, healthcare, public services in both India and the UK — will not let an autonomous agent operate on their code in a vendor's black box. They want the sandbox in an environment they govern, with their own logging and their own kill switch. Self-hosting and tunnelled control of agent runtimes is becoming a baseline expectation, a theme we explored in Anthropic's self-hosted sandboxes and MCP tunnels. The Ona deal is OpenAI answering the same demand from the other direction.
How to get ahead of it
You do not need OpenAI's blessing or Ona's product to start practising the skills this shift rewards. The patterns generalise across vendors. Concretely:
- Pick one repetitive, well-bounded task — a migration, a test-coverage push, a dependency bump across a monorepo — and write a brief precise enough that a stranger could execute it.
- Run it in a sandbox, not your main checkout. Constrain permissions deliberately and watch where the agent tries to overreach.
- Instrument the run so you can answer, afterwards, what it changed and what it cost. Long-running memory and state handling is its own discipline; our note on agent memory management patterns is a good primer.
- Review the output like a maintainer, not an author. Reject and re-brief rather than hand-fixing — that is the loop that scales.
- Publish what you learned. Building in public is now a hiring signal in its own right; see our playbook on building in public as an AI engineer.
Primary source: OpenAI's own announcement at openai.com/index/openai-to-acquire-ona. Coverage and additional detail from CNBC (11 June 2026), The Next Web and How2Shout.