Expo Wants to Rent Your Coding Agent a Phone, at Ten Cents a Minute

Expo shipped something on November 12 that sounds like plumbing and is not: cloud iOS simulators and Android emulators, sold to coding agents by the minute.
The pitch is narrow and specific. An agent that cannot run the app cannot see what the app does. On the web that gap closed years ago, because a headless browser is cheap and a screenshot is free. On mobile it stayed open, for a boring reason. iOS simulators only run on Apple hardware, cloud agents run on Linux, and Android emulators eat four gigabytes of RAM each before you have compiled anything.
So a developer running seven agents in parallel needs seven simulators, which means either a very large laptop or a queue. Expo's answer is to move the simulators into its cloud, hand each agent its own, and bill $0.10 per simulator minute.
What you actually get
A bug report filed in Linear goes to an agent. The agent reproduces it, writes a fix, runs it on a cloud simulator, and leaves a screen recording, screenshots, and logs on the pull request. Anyone on the team can open the running build from a link, with no TestFlight round trip.
That last part is the real product. TestFlight is a distribution channel with a review clock attached, and using it as an internal verification loop has always been awkward. A link that opens a live build is a different workflow, closer to how web teams review a preview deployment.
It works with Cursor, Claude Code, and Codex, and it runs on Linux, on Windows through WSL, or in a cloud sandbox. No Mac in the loop.
The price is the interesting part
Ten cents a minute is six dollars an hour, per simulator.

Run a ten minute verification pass per pull request and that is a dollar a PR. A team merging a hundred PRs a day is looking at a hundred dollars a day, or roughly three thousand a month, before anyone writes a line of product code. For a company already paying for model tokens, that is not a rounding error. It sits in the same budget conversation as inference.
The included allowance tells you what Expo expects. Starter, Production, and Enterprise plans each come with 60 minutes a month. Sixty minutes is six full verification runs. That is a trial, not a workflow. Expo knows it, which is why the pricing is usage-based rather than seat-based: they are betting that teams who try it will burn far more than the free tier once the loop becomes habitual.
Compare it to the alternative. Keeping a Mac mini in the office as a build host costs about a thousand dollars once and then a maintenance headache forever. Running seven of them costs seven thousand. At three thousand a month in simulator minutes, the cloud option is more expensive than owning hardware within three months, and cheaper than hiring the person who would maintain it within about a week.
Why this is the bottleneck nobody priced
For two years the story about coding agents has been about generation speed. Tokens per second, time to first patch, how many files an agent can hold in context. All of that got fast. What did not get fast is the loop after the patch exists.
On the web, the fix was trivial: run it in a browser, take a screenshot, look at it. On mobile, the equivalent step required hardware, and hardware does not scale with a git branch. Teams worked around it by giving agents only the parts of the codebase that could be unit tested, which left out everything that makes a mobile app a mobile app: navigation, layout on real screen sizes, keyboard behavior, permission prompts, and the specific way iOS renders a shadow that Android does not.
Expo's own framing is that the agent was the fast part. That is accurate and slightly self-serving, since Expo has been selling mobile infrastructure since 2015 and this is the missing piece of its own stack. But the observation holds: verification became the constraint the moment generation got cheap.
What a simulator will not tell you
A simulator reproduces the operating system, not the phone.
It will catch a broken layout, a crash on launch, a missing permission string, a navigation stack that loses its state. It will not catch a camera pipeline that fails on one vendor's image processing, a Bluetooth handshake that times out, thermal throttling after four minutes of use, or anything that depends on a specific modem. Those are the bugs that reach users, and they still need a device rack and a person holding the phone.
There is also a subtler risk in the workflow itself. A video of a fix running is evidence that the fix ran, not proof that it fixed the thing. Agents are very good at producing a passing state, and a recording of a passing state is exactly the artifact that gets attached to a pull request and skimmed by a reviewer. The format invites the same complacency that green CI badges did, and the answer is the same: someone has to look at the actual change.
The part that will get scrutinized
Expo says Agent Sessions is coming by the end of November: a secure layer that lets enterprises trigger agents from sources like customer feedback, with protection against prompt injection.
That is the right thing to be worried about. An agent that reads a customer bug report and then runs code on a simulator is an agent that takes instructions from strangers. The simulator is sandboxed, or should be, but the recording, the logs, and the build artifact all end up somewhere, and every one of those is an exfiltration path if the agent has been talked into using it.
Simulators add no model of their own, which helps. The blast radius is the simulator session, not the developer's machine. But the session has a network, and the session has a screen, and the screen gets recorded.
Who should try it this week
If your team runs one or two agents on a mobile codebase and reviews patches by hand, the free sixty minutes will tell you whether the loop feels right. Run a real bug, not a demo. Watch the recording. Decide whether a video of the fix running is more useful to you than a diff.
If you run agents at volume, the calculation is different and less comfortable. At a hundred PRs a day you are committing to a monthly bill that grows with your merge rate, which means your verification cost now scales with your team's productivity. That is a strange incentive to build into a toolchain, and it is the same incentive cloud CI created fifteen years ago. Everyone accepted it then. They will probably accept it now.
What is worth noting is how quickly the mobile toolchain has reorganized around agents. A year ago the question was whether an agent could write a React Native screen. This week the question is how many simulators you can afford to give it. The answer, apparently, is however many you are willing to pay ten cents a minute for.
Related articles
Six Amateurs, 1,113 Episodes, a Thousand Yuan of Compute an Episode: How an AI Drama Got Hand-Built
The first ten seasons of Armed Into Song run to roughly a thousand episodes and have passed ten billion cumulative views, with six core creators and not one of them film-school trained. Compute runs 1,000–2,000 yuan an episode; a normal shot takes three rolls, a battle twenty-plus, and anything that looks flat is graded a sixty and thrown out.
One server, one AI streamer: inside the code that runs a nightly livestream
open-yoki is the production code behind an AI host who streams nightly on Binance Square, with the persona, voice and money features stripped out. Headless Chrome draws the frame, ffmpeg pushes RTMP, a folder of Markdown decides what she feels and what she posts.
Pinky lets an agent drive your phone, and holds the Send button until you click Allow
Pinky connects any MCP agent to a real iPhone or Android phone and enforces approval in its own gateway rather than in a prompt. Banking apps are invisible, password fields are off limits, and on Android 13 the agent gets a hidden second screen.
ByteDance's SeedRealtime takes two frames a second and decides when to speak
Doubao-Seed-Realtime-1.0 folds camera frames, screen frames and speech into one full-duplex session, at no more than two frames per second. The interesting decisions are the frame rate, the proactive speaking, and the audit reply that ships switched on.