Multivergence Launches, Giving AI Agents State They Can Fork

Multivergence, the studio’s sixteenth project, is a small state engine for programs that need to try several futures before committing to one. An agent planning a trip can fork a branch per plan, give each a weight, attach constraints such as a fare limit and a deadline such as a seat hold, and feed in facts as they arrive. Branches that break a constraint or run out of time die on their own, the weights renormalise, and Promote makes the winning branch the real state in one step. Version 0.1.0 came out on 18 October as a Go library and a command with JSON output, with binaries for Linux, macOS and Windows. It’s pure Go, open source under Apache 2.0.

The engine never guesses a probability. The caller supplies weights and likelihoods, and the engine applies them by written rules, so the same calls give the same weights to the last bit. Branches share their data copy-on-write: one that changes 3 keys of a 10,000-key state costs about 3.9 KB where a full copy takes 2.1 MB. The store is one append-only file, and a promote is one record in it.

The test write-up kills writer processes with SIGKILL 1,000 times, 106 of them mid-promote, and every file reopens to one whole state. It also checks 86.8 million reads against a model that copies everything, and replays 300,000 random changes to the same weights. Where CodeInstances forks a recorded run at a step, Multivergence forks the state an agent is working on and keeps the branches apart until one wins.