Vons Releases a Compact, Host-Controlled Decision Component for AI Agent Workflows

Summary

Vons is a 0.1.0 research preview for compact local decision models used in frontier-agent workflows. A planner provides a state and a bounded candidate set, and Vons returns a structured choice or abstention; the host retains control over execution, policy, and consent. The research code includes a one-pass Direct scorer and a conditional Diffusion scorer using deterministic DDIM sampling. The project targets an English-first model asset bundle below 64 MiB, with CPU/WASM execution and optional WebGPU support, but says this is a target rather than a complete browser-package guarantee. The repository contains the Python contract, training, evaluation, export tools, a TypeScript SDK, tests, and a Chrome side-panel extension, while pretrained weights, tokenizer assets, raw benchmark data, private prompts, and experiment logs are distributed separately or withheld. The SDK can validate manifests and run supplied ONNX bundles, but a source checkout alone cannot perform pretrained inference, and no public npm package or hosted inference service is implied. The extension does not read pages, execute actions, or send prompts to a service; its version 0.1.0 Chrome Web Store submission was pending review on September 25, 2026. Vons reports synthetic pilot limitations, including no evidence of external-task generalization, separate browser and latency checks, and uncalibrated confidence values. The papers are under CC BY 4.0, while the source software uses a community and commercial license with commercial restrictions.