Laika Orbit, this website and its docs are built by Claude Code agents running inside Laika Orbit, with one person deciding what gets built and approving what ships. This is how that works, with examples from the week we launched.
In launch week, one set of chats renamed the product from its working name, another rebuilt the pricing page, another packaged the Mac app, and others wrote these docs, all on the same afternoon. Each chat works on its own branch in its own copy of the repo, so none overwrites another, and the fleet board shows which one needs a decision.
New pages go through a gauntlet. Agents build several versions, and critic agents compare each one against a page we admire (Linear's, for our pricing page) without being told which side is ours. Then the sides are swapped: a win in one position only is usually the position, so it doesn't count. A person looks at the survivors and picks.
Anything that takes more than a few minutes, like a build, a batch of screenshots or a site audit, reports to the runs page: its parts, how far along each is, when it should finish and what it produced. A job that goes quiet is flagged as stalled instead of looking busy. The audit that fixed this site's share images and phone layout before launch was one of these runs.
Builds and screenshot captures run on a Linux box under the desk, sent there by the same offload that ships in the app. The laptop stays quiet and cool while agents work, and the files come back when the job is done.
The repo's plans, decisions and notes are indexed by Laika Orbit recall, which every chat in the repo has as an MCP server. Asking "how do we sign the Mac app?" returns the one section that answers it in about a millisecond, with no model calls, instead of an agent opening a dozen files.
At the end of a day, away mode keeps the chats moving: the plainly safe steps are approved, stuck chats are brought back, and everything else waits. The next morning starts with one summary card of what finished and what needs a decision.
Decides what to build, reads the plans, and approves every push, deploy, domain change and anything that touches payments. Agents can't move saved passwords or stop running apps on their own: the safety checks stop them and hand the step to a person. The legal pages lost their draft labels only after our lawyer signed them off. And the public repo is exported from the private one with a scan that fails on personal names, home paths and keys, so private work never leaks out.
Everything on this page is in the free, open-source app.