One person directing many agents is the new unit of leverage. Loki is where that person describes the work, the agents build and ship it, and every change stays visible and reversible.
Agent capability has crossed the orchestration threshold. The bottleneck is no longer raw generation. It is human direction.
People are already running several agents at once across several projects — and people who never wrote code want the same result. Both need one place to direct, check and ship the work.
The winning architecture combines remote command with a clear choice of execution location. Loki supports the shared cloud builder for eligible accounts and Fleet Runner on your computer.
Open-weight models keep closing the gap on frontier capability. A layer that runs many vendors' agents is neutral to which model wins.
The same control patterns transfer to physical robotics. The market has not yet appreciated this.
A web workspace runs AI coding agents (Claude Code, Codex, Cursor, Antigravity, Grok) across many projects: from a plain description to a repository, a deployed site and a feedback loop that turns notes on that site into the next change. Invited accounts run on the shared cloud builder; Fleet Runner runs agents on the operator's own computer (Linux, Windows, Apple-silicon macOS). Every site deploy is checked on a phone-sized screen and rolled back if it got worse. Beta, in daily use.
We are raising to open the cloud builder to everyone who signs up, with each account isolated, to improve Fleet Runner, and to broaden open-model support.