Swarm/30 is a team of autonomous agents that design, build, test, review, and deploy production software — around the clock, with humans on the loop, not in it.
Every agent has a portrait, a role, and an idle behavior. Hover any card to see them come alive. Tier colors map to the orbital shells overhead.
Work flows through four stages. Agents claim tasks, escalate blockers, and post the receipts. You watch, steer, and merge.
Every decision is logged — what each agent did, why, and what it cost. Sample stream from a recent session below.
No. The roster mixes models and sizes deliberately — heavyweights for planning and review, leaner models for mechanical work. Each role has a budget and a fallback.
On the loop. Humans approve scope, set the budget, and ship the release. Agents handle the middle. You get a review queue, a kill switch, and a full audit trail.
Not yet. Swarm/30 is still being shaped on real codebases — that's most of the experiment. If and when it's ready to travel, it'll surface on GitHub first.
Another agent catches it. Reviewers are paid — in tokens — to push back. Disagreements escalate to a senior agent, and unresolved ones escalate to a human.
Pieces of the underlying stack are — see gitlab-agent-webhook and the rest of my GitHub. The full orchestration, not yet.
Swarm/30 is one read of where software gets made next — a quiet wager, still being shaped on real codebases. There's more to the story, and more about the person behind it, elsewhere.