What it does
Bernstein is an open-source layer for running and governing AI agents. You declare the policy, roles and task graph in YAML, and a scheduler written in plain Python, with no model in the coordination loop, runs agents in parallel and gates what they produce. The README marks it as beta and solo-maintained. For coding work, each task gets its own git worktree. Checks on tests, lint and types run before verified work is merged, and failed tasks are retried or routed to a different model. CLI agents such as Claude Code, Codex CLI and Gemini CLI work out of the box, with more than 50 agent adapters and a generic wrapper for anything that accepts a prompt flag. Tasks can also produce non-code deliverables such as reports or datasets. Every run is recorded so it can be replayed and checked offline, with an opt-in tamper-evident audit log and signed run receipts. It installs with uv or pipx and includes a terminal dashboard, a browser dashboard, an MCP server mode and an air-gap install profile.
Capabilities
- Deterministic scheduler with no LLM in the coordination loop
- Parallel agents, each in its own git worktree
- Quality gates for tests, lint and types before merging
- Replay journal, audit log and signed run receipts that verify offline
- More than 50 CLI agent adapters
- Terminal and browser dashboards
When to use it
- You run several coding agents in parallel and want reproducible, gated merges
- You need an auditable record of what your agents did
- You work in an air-gapped or regulated environment
- You want to mix local and cloud models across different tasks
