Last August, METR published a study that landed badly for the industry. Experienced developers using AI took 19% longer on their real work, while believing the entire time that they were moving faster. Those same engineers, on their own side projects, told the opposite story, sometimes getting more done in a weekend than they used to manage in a month. The person hadn't changed and neither had the model, only the setting around them had.

What slowed those teams down had little to do with the model itself and everything to do with the code around it. Reviews piled up and handoffs multiplied, and once anyone could generate code on demand, the real cost became making sense of all of it. Pull requests roughly doubled, and the hours spent reviewing them climbed to match.

Hand every engineer a rocket engine and you still don't have a rocket. You have a room full of thrust pointed in every direction, loud and expensive and arriving nowhere. The work worth doing is bolting those engines to one airframe so they all push the same way. That is what we built Covibes to do.

What it does today

Covibes is one workspace instead of a row of tools that don't talk to each other. The code, a running copy of the app, the review, and the agents doing the work are all in the same place.

You spawn agents and run several at once instead of supervising one and waiting on it. They can share a branch, or each can take its own branch with its own live preview when you want them isolated. Work that used to happen one step at a time now moves in parallel.

Those previews are live and shared. Anyone on the project can open the running app for any branch at a stable URL and watch it change as the agents work, not a screenshot someone pasted into a thread half an hour ago.

A lot of what we've built exists to stop agent output from quietly degrading the codebase, and it works as three layers stacked on each other. At the floor is Rox, our robustness engine. Before an agent's change reaches the code, it has to clear a set of deterministic checks that reach past style and lint into the architecture itself: that code depends on interfaces rather than concrete classes, that a decision has one owner instead of being re-derived in five places, that a bad migration can't hide behind a silent fallback or a compatibility shim. The checks can't be waived. There is no disabling a rule, ignoring a type error, or committing around it. When a new bad pattern slips through, we don't fix the one case; we add a rule that fails the whole category, so it can't come back. That floor is what lets a lot of agents and people commit all day without the codebase drifting, because it holds whether or not anyone is reviewing.

On top of Rox sit two more layers: Zeroshot, where independent validators review the work without knowing which agent produced it and send it back until it holds, and orchestration, which splits a job across many agents and coordinates them. Nothing in the stack is ever trusted to check its own work.

You can run it on your own terms: your own agents, in a browser or a desktop app that syncs both ways, offline when you need to, or entirely from the command line. Zeroshot, the orchestration layer, is open source under MIT today; Rox and the rest follow at launch. If we let it get worse, you can take the code and leave.

Why this is believable

We build Covibes with Covibes. Since December our repositories have absorbed the work below, and we are not planning to hire engineers to sustain it.

Zeroshot sits at 1,444 stars and 125 forks, starred by engineers at Microsoft, AWS, Google, Apple, Cloudflare, Shopify, Atlassian, and Adobe. Something that can build a multi-agent development platform can build whatever you have in front of you.

Where it is going

The engineering loop is the part that works today: agents, code, preview, and review in one place. The planning, docs, and issue tracking around it are still basic, and building those out in the same workspace is what comes next. For now, judge Covibes on the loop that already runs.

Come build something

If you want to try it on something real you're working on, tell me and I'll get you set up this week.