# Who owns the coding-agent runtime?

URL: https://www.thedeepfeed.ai/posts/2026-06-13-who-owns-the-coding-agent-runtime/
Category: Agents
Published: 2026-06-13
Author: the-deep-feed
Tags: ai-agents, agent-runtime, sandboxes, openai, codex, niteshift
Kind: deep

> In 48 hours the question of where long-running coding agents actually run got both answers at once: an indie $7M bet against lock-in, and a frontier-lab acquisition that pulls the runtime layer inside Codex.

## TL;DR

- On **June 10**, two ex-**Datadog** engineers launched **Niteshift** with **$7M** led by **Greylock**, pitching a neutral runtime for any coding agent and an explicit bet against frontier-lab lock-in.
- On **June 11**, **OpenAI** said it would acquire **Ona** (the company formerly called Gitpod) to fold secure, customer-controlled cloud execution into **Codex**. The runtime layer's first frontier-lab capture.
- The bottleneck moved from the model to the **environment** where the agent runs, verifies, and ships. Both camps noticed in the same 48-hour window.
- The map splits three ways by **who owns the agent's lifecycle**: neutral SDK primitives, full-stack clouds, and now a lab-captured tier. The open question is whether neutral environments hold once the labs start buying.

The interesting fight in AI coding is no longer about which model writes the better function. For most of the work that ships, the frontier models are close enough that the difference is rounding error. The fight that decides who captures the value has moved one layer down, to a duller-sounding question: where does the agent actually *run*? Where does it install dependencies, execute the test suite, hit a real database, fail, retry, and open a pull request a human can merge?

In a 48-hour window in June 2026, that question got two answers at once, from opposite ends of the market. On June 10, two former early **Datadog** engineers launched a startup called **Niteshift** with a $7 million seed round and a thesis that the runtime should stay neutral. The next morning, **OpenAI** announced it was buying a company called **Ona** to pull the runtime inside Codex. One bet says the environment must serve every agent. The other says the most valuable environment is the one your lab controls end to end. Both cannot be right, and the gap between them is the most important unsettled question in agent infrastructure.

![Two hands reaching for the same small glowing cube from opposite sides of a wide cream field, one open palm and one closing grip, a single red edge on the contested object](/post-images/2026-06-13-who-owns-the-coding-agent-runtime/two-hands-one-runtime.jpg)

# The bottleneck nobody was pricing

Start with the claim both sides actually agree on, because the agreement is the news.

For two years the binding constraint on coding agents was the model. A weak model wrote plausible code that did not compile; a strong model wrote code that did. As the models climbed, that constraint loosened, and a different one became visible. An agent that can write a browser engine or port a runtime from one language to another in a week, as Niteshift's [launch post](https://niteshift.dev/blog/introducing-niteshift) catalogs, still cannot reliably *ship* that work, because shipping requires a place to run it: a filesystem, package managers, a network, secrets, a way to execute the tests and read the failures and try again. On a developer's laptop, that place exists but dies the moment the lid closes. In CI, it exists but is built for deterministic pipelines, not for an agent that wants to poke around for an hour.

OpenAI described the same gap in plainer terms when it announced the Ona deal. Codex now has [more than 5 million weekly users, up 400%](https://openai.com/index/openai-to-acquire-ona/) from earlier in the year, and the company's stated reason for the acquisition was to give those agents "secure, customer-controlled cloud infrastructure for long-running agents." [SiliconANGLE](https://siliconangle.com/2026/06/11/openai-acquires-ai-agent-orchestration-startup-ona/) put the operational version of the problem bluntly: developers run agents on local machines, and "when a workstation is turned off, the AI agents that it runs stop too." The whole pitch for a runtime layer is the removal of that sentence.

So the diagnosis is shared across a $7 million seed and a frontier lab's M&A budget: the environment, not the model, is now what gates how much real work an agent finishes. What the two camps disagree about is who should own it.

# The indie bet: the runtime stays neutral

Niteshift's argument is the cleaner one to state. The company, founded by former Datadog engineers Sajid Mehmood and Conor Branagan, launched general availability and the funding round on the same day, June 10, with [$7 million led by Greylock](https://greylock.com/portfolio-news/introducing-niteshift-the-full-stack-cloud-for-coding-agents/) and participation from Amplify Partners, BoxGroup, and SV Angel. [TechCrunch](https://techcrunch.com/2026/06/10/datadog-veterans-launch-ai-coding-startup-niteshift-on-a-bet-against-big-ai-lock-in/) framed the whole thing as "a bet against Big AI lock-in," and that framing is the product.

Niteshift describes itself as "the full-stack cloud for coding agents," and the operative word is *agents*, plural and unowned. It gives **Claude Code**, OpenAI's Codex, and open-source agents the same thing: a real environment to run in, verify against, and ship from. You hand it a prompt, a bug report, a Linear ticket, or a half-finished pull request, and it returns a pull request with evidence the change works. The neutrality is not incidental. It is the entire reason to exist. If your company's value is in being the environment that any model can plug into, then the day a frontier lab can lock its model to its own environment is the day your value evaporates. Niteshift is selling insurance against exactly that day.

The investor logic rhymes with the founder logic. Greylock's Jerry Chen led the round, and the supporting cast on the cap table includes Datadog's founders Olivier Pomel and Alexis Le-Quoc, plus Reid Hoffman, names that read as a wager that enterprises will refuse to let one lab own both the brain and the body of their automation. The enterprise objection writes itself: a company that has standardized on Codex this quarter does not want its entire shipping pipeline to become un-portable the moment a better model ships somewhere else. A neutral runtime is the hedge that keeps the model swappable.

That is the indie thesis in one line. The environment is the durable asset precisely because models are not, and the environment is only durable if it refuses to belong to any single lab.

# The lab bet: the runtime comes inside

OpenAI's move is the exact negation, executed one day later.

On June 11, OpenAI said it had agreed to acquire Ona, with terms undisclosed and the deal subject to regulatory approval; [Bloomberg reported it first](https://www.trendingtopics.eu/openai-buys-german-ai-startup-ona-to-strenghten-codex/), and OpenAI and Ona confirmed it in matching posts. Ona's technology is secure cloud execution and orchestration for long-running agents, the persistent, customer-controlled environment that a Codex session can keep working in for hours or days without a laptop staying awake. [The Decoder](https://the-decoder.com/openai-buys-ona-to-push-codex-toward-long-running-autonomous-coding-tasks/) summarized the intent as pushing Codex "toward long-running, autonomous coding tasks." In the language of this piece: OpenAI just bought the body to go with its brain.

The traction numbers on both sides explain the urgency. OpenAI cited that 400% jump in weekly Codex users. Ona, in its [own announcement](https://ona.com/stories/ona-joins-openai), said weekly agent sessions on its platform had grown 13x since January. Two curves bending up at the same time, pointed at the same workload, is a strong reason to merge them rather than let a third party sit in the middle collecting rent on the connection. Vertical integration is rarely about elegance. It is about refusing to share the margin on a layer you think is about to matter.

For Ona, the deal is also the end of a long identity migration, and that migration is the investigative thread worth pulling.

# The Ona thread: Gitpod, renamed, then exited

Ona did not start as an agent-runtime company. It started as **Gitpod**, founded in Kiel, Germany, in 2020, building cloud development environments, the "spin up a ready-to-code workspace in your browser" product that a generation of developers used to escape laptop setup hell. The cloud-dev-environment market got crowded and commoditized, and the company did what companies in commoditizing markets do: it found the next curve. It rebranded as Ona and repositioned the same core asset, secure ephemeral cloud environments, as the substrate for autonomous agents rather than human developers. [The Decoder](https://the-decoder.com/openai-buys-ona-to-push-codex-toward-long-running-autonomous-coding-tasks/) traced the lineage plainly: Ona, "previously known as Gitpod," founded in Kiel in 2020.

Ona's founder framed the exit as continuation rather than ending, and the framing matters because it is also the sales pitch for what capture buys a runtime company: distribution it could never have built alone.

> I always thought selling the company would feel like an ending. Instead, our life's work is getting bigger and more important.
>
> — Johannes Landgraf, CEO of Ona, [Ona is joining OpenAI](https://ona.com/stories/ona-joins-openai), Jun 11, 2026

The pivot worked well enough to attract a buyer at the exact moment the runtime question went hot. And the buyer is a US frontier lab, which makes Ona the second European AI company in roughly three weeks to be absorbed into a frontier lab's stack. [Trending Topics](https://www.trendingtopics.eu/openai-buys-german-ai-startup-ona-to-strenghten-codex/) noted the pattern in the same breath: Ona and the Linz-founded **Emmi AI**, which [Mistral acquired in May](https://www.emmi.ai/news/mistral-ai-acquires-emmi-ai), were both Speedinvest portfolio companies, exiting to frontier labs weeks apart. Emmi was physics AI for industrial engineering, a different domain entirely, so this is not a runtime-specific drain. It is a broader pattern: European deep-tech startups that build something a frontier lab needs are getting bought by frontier labs before they get the chance to become frontier labs themselves. Ona is the version of that story where the thing being absorbed is the agent's runtime.

Which is what makes the timing with Niteshift feel less like coincidence and more like a fork in the road being drawn in real time. The same week one company raised money to keep the runtime neutral, another sold the runtime to the largest model vendor on earth.

# The map, split by who owns the lifecycle

The reason this matters beyond two news items is that the agent-runtime layer is now a real category with real money in it, and the two June deals are not outliers; they are the two newest entries in a field that has been filling up for a year. The honest way to read the field is not open-source versus closed, or fast cold-start versus slow. It is *who owns the agent's lifecycle*: who controls the environment from the moment the agent wakes up to the moment its work merges.

Three tiers fall out of that question.

| Tier | Who owns the lifecycle | Players | Posture | Funding signal |
|---|---|---|---|---|
| **SDK primitive** | The builder. You get a sandbox API; you orchestrate the agent yourself. | E2B, Modal, Daytona, Runloop, Blaxel | Model-agnostic by construction; sells compute + isolation | Modal $355M @ $4.65B; E2B $21M Series A; Daytona $24M Series A |
| **Full-stack cloud** | The platform. The environment ships bundled with a deploy/edge story. | Vercel Sandbox, Cloudflare Sandboxes, AWS AgentCore, Google Agent Sandbox | Runtime as a feature of a larger cloud relationship | Vercel Series F $300M @ $9.3B; Cloudflare, AWS, Google are public/parent-funded |
| **Lab-captured** | The frontier lab. The environment is wired to one model family. | Ona (→ OpenAI / Codex) | Vertical integration; runtime tuned to the lab's agent | Acquisition; terms undisclosed |
| **Neutral runtime (contested)** | Explicitly the builder, against the lab. | Niteshift, E2B | Sells portability as the product | Niteshift $7M seed (Greylock) |

The tiers are not cleanly exclusive, and that is the point. **E2B** appears twice because it is both an SDK primitive and, increasingly, a standard-bearer for the neutrality argument; its whole pitch is that an open, model-agnostic sandbox serves every agent. **Modal**, the richest player in the layer at a [$4.65 billion valuation](https://modal.com/blog/modal-series-c) after its $355 million Series C, sells raw runtime to anyone, including the labs, and reports more than $300 million in ARR doing it. **Daytona** rebuilt itself from a dev-environment manager into an agent-runtime company, the same pivot Ona made, and has not been bought. **Vercel**, **Cloudflare**, **AWS**, and **Google** sit in the full-stack tier because for them the sandbox is a feature that deepens a cloud relationship, not a standalone business; the runtime is bait for the larger hook.

What June added is the fourth row and the tension between it and the bottom row. Until this month, no frontier lab owned a runtime outright. Ona changes that, and Niteshift exists to argue it should not become the norm.

![Three stacked horizontal tiers across a cream field rendered as layered strata, the lowest tier broad and even, the top tier fractured and pulled toward one corner where a single red marker sits](/post-images/2026-06-13-who-owns-the-coding-agent-runtime/three-tier-strata.jpg)

# Why the labs want the body, not just the brain

The vertical-integration instinct is easy to caricature as land-grabbing, but there is a real engineering reason a lab wants to own the runtime, and it is worth stating fairly.

A long-running agent is a feedback loop, not a single inference. It proposes an edit, runs the code, reads the error, and revises. The quality of that loop depends heavily on how tightly the model is coupled to the environment: how fast the sandbox boots, how cleanly the test output streams back, how the agent's tool calls map to real filesystem and network operations, how failures are surfaced in a form the model can act on. When the lab owns both ends, it can co-design them. It can tune Codex's behavior to Ona's execution semantics and tune Ona's execution to what Codex needs to see. A neutral runtime, serving every agent equally, cannot specialize that hard for any one of them without breaking its neutrality.

![A closed loop drawn as a continuous line cycling between a stylized brain shape and a gear-like machine block, the loop tightening where the two meet, one red segment marking the join](/post-images/2026-06-13-who-owns-the-coding-agent-runtime/agent-loop-join.jpg)

That is the strongest case for capture. The labs are betting that the agent loop has enough surface area between model and machine that owning both halves produces a meaningfully better product than two best-of-breed halves bolted together at an API boundary. If they are right, the lab-captured tier wins the high end, and the neutral runtimes get pushed toward the long tail of teams that value portability over peak performance.

The counter-bet, the one Niteshift and E2B are making, is that the coupling gains are real but smaller than they look, and that they are swamped by a market force pulling the other way: enterprises do not want to be locked in. A buyer who standardizes on a lab-owned runtime has handed that lab a switch, the same kind of switch that, in an adjacent corner of this industry, [a single government letter just proved can go dark without warning](https://thedeepfeed.com/posts/2026-06-14-export-control-frontier-model-shutoff/). The neutral camp is selling the promise that your shipping pipeline survives a change of model, a change of vendor, or a change of policy. For a CIO who has watched any of those things happen, that promise is not abstract.

# What the OSS runtimes already taught us

This is not the first time the agent-infrastructure stack has split along an ownership axis, and the earlier split is a useful tell.

In the open-source agent-framework layer, the recurring pattern was that the vendor SDKs from the big labs kept [absorbing the primitives that independent frameworks had built](https://thedeepfeed.com/posts/2026-06-02-oss-agent-runtimes-five-wheels/), memory, sandboxing, tool governance, into their hosted platforms. The independents solved the hard problems first; the platforms folded that work in and bundled it with distribution. The runtime layer is now rhyming with that. Ona spent years building secure ephemeral environments as an independent; OpenAI is folding that capability into Codex and bundling it with the largest agent install base in the market. The pattern is consistent: labs do not need to invent the runtime, they need to acquire it and attach it to distribution they already own.

If the framework layer is the precedent, the neutral runtimes have a narrow but real path. The OSS frameworks that survived absorption did so by being genuinely better at one hard thing, or by being the portability layer enterprises insisted on as a hedge. Niteshift is explicitly going for the second. Whether that holds depends on a variable no one can read yet: how badly enterprises fear lock-in once the lab-owned runtimes are simply faster and more polished because the lab co-designed both halves.

# The question the next 18 months answer

Two deals, one week, opposite directions. That is the shape of the moment, and it is genuinely undecided, which is rare enough in this industry to say plainly.

Here is the test that resolves it. Watch whether the other frontier labs follow OpenAI into the runtime layer. If Anthropic and Google move to own or tightly bind their own execution environments for Claude Code and Gemini's agents, then capture is the trend, vertical integration wins the high end, and the neutral runtimes are fighting for the hedge-buyers and the long tail. If instead the labs keep renting from neutral providers, if Codex remains an outlier and the rest of the market routes through E2B, Modal, and the full-stack clouds, then neutrality held, and Ona was a one-off rather than a template.

My read is that the runtime is too strategic to stay rented for the labs that can afford to buy. The agent loop is where usage, lock-in, and the next round of differentiation all converge, and a lab that owns the model but rents the environment is leaving its best loop half-built. Expect more captures. But that same logic is exactly why the neutral bet is fundable rather than naive: the harder the labs pull the runtime inside, the more an enterprise that refuses single-vendor capture will pay for the one provider that promises its pipeline cannot be switched off by someone else's letter. The runtime layer is not consolidating toward one answer. It is splitting into two markets, the lab-owned high end and the neutral hedge, and the June deals were the first clean shot of each.

## Sources

- [TechCrunch — Datadog veterans launch AI coding startup Niteshift](https://techcrunch.com/2026/06/10/datadog-veterans-launch-ai-coding-startup-niteshift-on-a-bet-against-big-ai-lock-in/)
- [Greylock — Introducing Niteshift: the full-stack cloud for coding agents](https://greylock.com/portfolio-news/introducing-niteshift-the-full-stack-cloud-for-coding-agents/)
- [Niteshift — Introducing Niteshift (launch blog)](https://niteshift.dev/blog/introducing-niteshift)
- [OpenAI — OpenAI to acquire Ona](https://openai.com/index/openai-to-acquire-ona/)
- [Ona — Ona is joining OpenAI](https://ona.com/stories/ona-joins-openai)
- [CNBC — OpenAI to acquire Ona to support Codex](https://www.cnbc.com/2026/06/11/open-ai-ona-acquisition-codex.html)
- [SiliconANGLE — OpenAI acquires AI agent orchestration startup Ona](https://siliconangle.com/2026/06/11/openai-acquires-ai-agent-orchestration-startup-ona/)
- [The Decoder — OpenAI buys Ona to push Codex toward long-running tasks](https://the-decoder.com/openai-buys-ona-to-push-codex-toward-long-running-autonomous-coding-tasks/)
- [Trending Topics — OpenAI buys German AI startup Ona to strengthen Codex](https://www.trendingtopics.eu/openai-buys-german-ai-startup-ona-to-strenghten-codex/)
- [Modal — Series C announcement](https://modal.com/blog/modal-series-c)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-06-13-who-owns-the-coding-agent-runtime/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt