I ran 12 AI agents for 47 hours straight. Here's what broke (and what shipped).
Hey Product Hunt! I'm Alex, maker of Bernstein.
Last month I ran 12 AI coding agents on a single laptop for 47 hours - Claude Code, Codex CLI, Gemini CLI working in parallel. 737 tickets closed. 826 commits. 15.7 tickets per hour.
Everything that could go wrong did: merge conflicts, context drift, runaway token costs, agents overwriting each other's work, worktrees missing node_modules, port collisions, agents stuck in infinite loops burning through API credits.
Bernstein is every fix I built during that sprint, packaged into an open-source tool:
- Deterministic orchestrator — zero LLM tokens on coordination, pure Python
- Per-run cost budgets — hard-stop agents when the dollar cap is reached
- Parallel git worktrees — each agent gets its own isolated workspace
- Built-in janitor — verifies tests pass before anything gets committed
- Mix models in one run — cheap agents for boilerplate, heavy models for architecture
No framework to learn. No vendor lock-in.
pip install bernsteinand go.
What would YOU want from a multi-agent coding orchestrator?
Drop your wishlist below - I'm building in public and shipping fast.
GitHub: https://github.com/chernistry/bernstein
Docs: https://chernistry.github.io/bernstein/
Replies