Bernstein – open-source governance layer for AI agents

Bernstein is an open-source governance and orchestration layer for running multiple AI coding agents in parallel. Its deterministic Python scheduler assigns work without using a language model to decide what runs, what is blocked, or what gets merged, allowing plans to replay consistently. Agents operate in separate Git worktrees; Bernstein checks each proposed change against configured tests, linting, and type checks, and merges only verified diffs. It supports more than 50 command-line agent adapters, including Claude Code, Codex, and Gemini CLI, and can also apply artifact contracts to research, datasets, and audit packs. The project is Apache 2.0 software, currently in beta and built by a single maintainer.

A central claim is that agent work can be independently verified rather than trusted on the strength of a dashboard or report. Bernstein records run events in a local, hash-linked audit log; with a configured signing key, it also creates an Ed25519-signed receipt that can be checked offline. Telemetry is local by default, with optional export to monitoring systems, and deployments can run on a laptop, behind a firewall, in Docker, on Cloudflare Workers, or across Kubernetes. Some features, such as model-assisted agent selection and judging, are optional; the scheduler itself remains deterministic. Bernstein is intended to govern and verify agents, not replace them, and can use Claude Code or other tools as its underlying workers.