Plan. Build. Test. Review.
On your own hardware.

Delegatron started in April 2026 with one question: could ordinary local models — the kind that fit on a single gaming GPU — write code you could actually rely on, if they had to check each other's work?

A single model asked to write code will cheerfully say it's done. So Delegatron splits the job the way a small team would. An architect writes the plan and the acceptance criteria. An implementer writes the code. A sandbox runs the tests — the one opinion nobody can argue with. A reviewer reads the change and the evidence, then decides. When something fails, the loop goes around again with a clean slate and the failure in hand.

Most of it was built in one intense month by a single developer at StoriaBrevi, much of it pair-programmed with an AI coding assistant — which felt right for a tool about delegating work to AI. It was revisited in the autumn to bring it up to date with newer models.

It began as a possible product. In the end we decided it's more useful as an open, readable example of how a careful multi-agent pipeline fits together: the state machine, the evidence rules, the sandboxing and the budget limits. So it's published under the AGPL-3.0 — anyone can read it, run it and learn from it, and anything built on it as a network service has to stay open too.