Mission
We are building an agent-first cloud: somewhere to run software where the agent that wrote it can also ship it.
An agent can write a service, test it, and fix its own bug. Then it stops, because the next step is a browser.
That is the mission above, and removing that stop is the whole of it. Everything below is why we hold it.
The gap
Engineering changed on my team faster than our infrastructure did. We now write purpose-built software for almost every problem we touch: migration scripts, data repair tools, one-off pipelines, validation scripts that run twice and die. Over a hundred one-shot programs in a short period. Before ChatGPT that would have been extravagant. Custom software for one customer migration cost half a week of attention, so you did it in SQL, in a spreadsheet, or you didn't do it.
Almost none of that software has a proper home. It runs on laptops against production data. In shells opened on production pods. With credentials wherever they happened to land. Not because we don't know better, but because the only official path to real infrastructure was built for production services, and nobody pushes a 200-line run-twice script through repo, CI, registry and IaC. The ceremony lost, so the hygiene lost with it.
Writing the software takes twenty minutes. Running it reliably takes the rest of the day.
- Write · ~20 min
- Deploy · rest of day
AI collapsed the cost of creating software. The cost of running it properly barely moved. Every team adopting agents is opening this same gap right now, and the gap is where the risk pools: production credentials on personal machines, irreversible scripts with no audit trail, work that exists only in one person's terminal history.
One cloud, the whole range
The smallest thing worth deploying is a single function. One handler, one script, one job that runs twice and dies. The largest is an orchestra of small services that call each other.
Most clouds make you pick a different product for each shape. A functions product, a containers product, a queue product, each with its own console, its own limits, and its own way of describing a build. You choose between them before you have learned anything about the problem, and changing your mind later means starting over in a different console.
One definition and one command should cover the whole span. The shape of the software is a detail of the software, not a purchasing decision.
The bar is a Google Doc
Making a Google Doc costs nothing first. No project to create, no region to choose, no billing account to attach, no framework preset to confirm. It exists, and it has a URL.
Deploying should feel like that.
Sharing should too, and this is the part hosting usually gets wrong. A Doc is private until you name someone. Then that person can open it, and nobody else can. No login page to build, no auth provider to wire up, no token to pass around, no decision about whether this thing is worth protecting. A running application deserves the same default.
What we believe
Small software deserves real infrastructure. Not production ceremony scaled down. Infrastructure shaped for its actual life: born in a chat session, run intermittently, asleep most of the time.
An agent should have a computer of its own. Durable, up in seconds, owned by no model provider. The agents will keep changing. The computer shouldn't have to.
Idle should cost almost nothing. A thousand sleeping programs should cost about what a thousand files cost. If keeping a tool alive costs real money, people kill their tools, and then rebuild them worse.
Secure should be the default, not the reward for effort. Credentials belong on the machine where the work runs, scoped to that work, never on a laptop. The hygiene has to come free with the path of least resistance, because the last three years proved it won't come any other way.
Where we are
Deploying is the part that is close: a SHED file in the repository, two commands, and the build runs on Shed rather than on your laptop.
The rest is not built. A SHED file describes one application today, not several. Private sharing is not available, and neither is sign-in in front of your app. Both are on the home page marked as coming, which is exactly where they stand. We would rather name the distance than write a mission that quietly implies it closed.
Where this goes
The near picture: you leave your agent running on a remote machine, the migration finishes at 2am, you reopen the same session from your phone in the morning, and the dashboard it wrote is still up.
The far one is bigger. When running software becomes as cheap as writing it just became, software stops being rationed. Every recurring annoyance in a company becomes a small program someone spun up in an afternoon. The hundred scripts my team wrote this quarter are an early trickle of the largest volume of software ever created, and it will all need somewhere to live.
This is for people, and agents, shipping small software, from one function to a handful of services that talk to each other. It is not a replacement for your production estate.
Every piece of this exists somewhere already. Nobody has put them in one place for the software that's actually being written now.
That's the gap. Shed closes it. If it's your gap too, follow along.