What it does
Bernstein schedules 40+ CLI coding agents (Claude Code, Codex, Gemini CLI, and others) to work on a single goal in parallel, each in an isolated git worktree. The scheduler is deterministic—pure Python logic, no LLM in the coordination loop—so you can replay yesterday's plan and get yesterday's task graph. Every scheduling decision is HMAC-signed into an audit chain; every artifact write tracks its producer, inputs, prompt hash, model, and cost. The orchestrator serializes merges to prevent race conditions when multiple agents complete simultaneously.
Who it's for
Teams running three or more CLI coding agents and burning $1k+/month on agent calls. Operators in compliance-sensitive workflows who need tamper-evident audit trails and want credentials to stay in their environment, not in a SaaS hop. Platform teams rolling agents into client repos without storing external secrets.
Common use cases
- Run 3+ agents on parallel worktrees against a single coding goal, merge serially without race conditions
- Audit agent decisions: grep the plaintext HMAC-signed log to see what the orchestrator decided and why
- Replay a previous run to verify determinism or reduce redundant model calls
- Track artifact lineage: which agent produced each file, with what prompt, from what inputs, and at what cost
- Drop agents into client repositories without storing API keys externally
Setup pitfalls
- 173 secrets detected in codebase—review the security model in
docs/security/before deploying in compliance contexts - Requires Python 3.12+; verify compatibility with your agent CLIs before spawning
- Filesystem write access required for worktree creation, audit logs, and artifact storage; plan disk space and consider sandboxing constraints
- High risk classification—bearer-token auth and JWS agent cards have security implications; read the manager auth and audit log documentation before use in untrusted environments