Image: GitHub
Why it mattersMost multi-agent tools trust a bot's own word that a task is done. Kwant only marks a task verified after it runs the check.
Running several AI agents at once is easy. Checking that any one of them actually did the work is not. Every multi-agent tool asks the next bot to trust what the last one said.
Alex Sergivan released kwant, a self-hosted platform that puts the check in a different place. Kwant runs the test itself, reads the output and then marks the task done. The repository is on GitHub with 440 stars in 30 days, 36 forks, and an AGPL-3.0 licence. It is written in Rust with an embedded React front end.
What kwant does
Kwant gives a user a set of named bots that share chat spaces. A user can mention one bot and it will hand work to another. A bot can be scheduled. A bot can keep working while the user is away.
Each task appears as a card on a Mission board. A card moves to Verified only when kwant has run its own check on the output. The check is something kwant can see, such as a test or a file check. It is not a chat message from the bot that claims the work is done. The card keeps the receipt.
That matters. In most multi-agent tools, a bot writes "done" and the next bot picks up on faith. When the first bot got the work wrong, the second bot works on something broken. The verify step catches the break before the handoff.
What a user needs
Kwant runs on macOS, Linux or Windows. The data stays in a folder on the user's machine called ~/.kwnt. Nothing is sent to the creator.
The platform connects to accounts the user already has. It supports Claude, Codex and Grok through their CLI logins. It also takes API keys from Anthropic, OpenAI or xAI. A user can point it at any OpenAI-compatible or Anthropic-compatible server.
Dev mode and routines
Kwant has a Dev mode for teams that write code. The Dev desk shows what landed in each team's repository in plain words. A bot that did not write a pull request reviews it. Tests and secrets are included in that review.
A user can ask for a routine such as "every morning" and get one scheduled. Each routine gets Undo, Edit and Pause buttons.
Where this fits
Kwant sits between a single coding agent and a full service platform. A founder or a small team may run three or four accounts across Claude, Codex and Grok. Kwant puts those agents in one window with one set of rules. The verify step is the part worth testing. An agent that writes a test and runs it against its own code still fails on a wrong test. The useful check is whether kwant's test picks up a known fault the user puts in on purpose.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
