bernstein

Bernstein takes a goal, splits the work across several AI coding agents (Claude Code, Codex, Gemini CLI and many more), runs tests and merges the code that passes. The README says it has 31 CLI agent adapters, 30 third-party wrappers plus a generic wrapper for any tool with a `--prompt` flag, and that you can mix agents in one run, for example cheap local models for boilerplate and heavier cloud models for architecture. Any adapter can also act as the internal scheduler LLM.

Its stated difference is that scheduling is not done by an LLM. It makes one LLM call to decompose the goal into tasks, and everything after that (running agents in parallel, isolating their branches, running tests, routing retries) is plain Python, so runs are reproducible and logged. The flow is decompose, spawn each agent in its own git worktree, verify with concrete signals such as tests, files, lint and types, then merge verified work into main. Failed tasks are retried or sent to a different model.

Install with `curl -fsSL https://bernstein.run/install.sh | sh` on macOS or Linux, a PowerShell one-liner on Windows, or `pip`, `pipx`, `uv`, `brew` or `npx`. Then `bernstein init` and `bernstein -g "Add JWT auth ..."`; `--headless` gives JSON output for CI, `bernstein run --dry-run plan.yaml` previews tasks and estimated cost, and `bernstein live` shows a dashboard. It is Apache-2.0 licensed and was last pushed on 2026-10-12. The README links a known-limitations page. This review is based on the README and repository data; we have not run it ourselves.