Maelstrom Lab is an experiment with a simple premise: one engineer should not be able to run a portfolio of real products — Leverage, Nora AI, wave.dance, and more. A 24-hour day doesn’t allow it. It works anyway, because most of the work isn’t done by a human anymore.
The setup
The studio runs on a fleet of AI agents working inside one monorepo, under hard rules:
- Agents work on short-lived branches.
mainis untouchable, gates must be green, and every PR gets reviewed by the human. - Each agent has a defined role and boundary — research, build, design, content, monitoring, support. The roster on the homepage is not a metaphor; those are the actual lanes work gets routed through.
- The human is the bottleneck for exactly two things: taste and trust. Everything else is delegated or automated.
Measuring whether an agent setup like this actually behaves is its own problem — that’s what agentsgym came out of.
Proof of work
This site was built the same way it describes: an agent read the old code, wrote the plan, split the company site from the founder’s personal site, and shipped the redesign — including the canvas engine from wave.dance rendering the background of this page. The human reviewed the diff.
What the ship log is for
Build logs from the products, agent workflows that worked (and the ones that failed), and the numbers behind them. The polished pitch is the homepage; this is the record of what actually shipped.