sipyourdrink-ltd/bernstein

What it solves

Bernstein provides a governance and coordination layer for AI agents, addressing the lack of reproducibility, auditability, and safety when running multiple agents. It prevents "flaky" agent behavior by removing the LLM from the coordination loop and ensures that agent actions are verifiable and isolated, preventing them from corrupting a shared workspace.

How it works

Bernstein uses a deterministic scheduler written in plain Python to manage agent workloads based on "policy as code."

1. Decomposition: A manager breaks a goal into specific tasks with assigned roles and completion signals.
2. Isolation: Each coding task is spawned in its own git worktree (or a plain directory for non-code artifacts), ensuring agents do not share mutable state.
3. Verification: A "janitor" process verifies concrete signals, such as passing tests or linting, before work is merged.
4. Auditability: The system maintains a replay journal and a lineage spine. When audit mode is enabled, it creates an HMAC-chained audit log and signed receipts that can be verified offline to prove exactly what happened during a run.

Who it’s for

It is designed for developers and organizations that need to run AI agents (such as CLI coding agents) in a production-like environment where security, compliance, and verifiable evidence of work are required.

Highlights

• Deterministic Coordination: No LLM is used for scheduling, making runs reproducible.
• Offline Verification: Signed receipts allow reviewers to verify the integrity of a run without needing the original environment.
• Broad Compatibility: Includes adapters for over 40 CLI agents (e.g., Claude Code, Aider, Gemini CLI).
• Isolated Workspaces: Uses git worktrees to keep the main branch clean and isolate agent tasks.
• Policy-as-Code: Uses YAML manifests to define phases, roles, and dependencies for complex agent workflows.