Bernstein: Open-Source AI Agent Governance Layer Released

What changed

Bernstein, an open-source governance layer for AI agents, has been released. This project, maintained by sipyourdrink-ltd, operates on a "policy as code" principle, enabling users to define rules for AI agent actions declaratively. These rules dictate aspects such as who can perform specific tasks, what requires approval, and what actions must be recorded. Bernstein then enforces these policies and generates a verifiable, replayable record of all operations. A key feature is its deterministic scheduler, which uses plain Python for coordination, eliminating reliance on AI models in the decision-making loop. This ensures that runs are reproducible end-to-end; executing the same plan twice will yield the same task graph. The system supports over 50 CLI coding agents out-of-the-box, including Claude Code, Codex, and Gemini CLI, and is designed to govern any agent workload, producing deliverables like code diffs, research reports, datasets, or audit evidence packs. Bernstein also includes an air-gap install profile for secure, offline deployment. The latest release is v3.24.1.

Key differentiating features highlighted include:

• No LLM in coordination: Scheduling is handled by plain Python, ensuring reproducibility.
• Hardware key evidence: Supports offline PIV custody and FIDO2 approval verifiers for secure receipt protocols.
• Post-run verification: A replay journal and an always-on lineage spine record every step, with an optional HMAC-chained audit log for offline verification of receipts. Non-determinism is surfaced as a hash mismatch at the exact step.
• Isolation by construction: Coding tasks run in isolated git worktrees behind merge gates, and artifact-mode tasks use dedicated workspaces. Agents do not share mutable workspaces by default, with the task backlog being the only shared atomic state.
• Broad compatibility and local operation: Supports over 50 CLI agent adapters and a generic --prompt wrapper, with no SaaS hop or third-party data plane dependency.

The system allows defining runs via a YAML file, specifying phases, roles, and dependencies. Each node in the execution graph is claimed by an agent whose role is permitted by the phase. Deliverables can be code, research, operational changes, or a mix, with artifact contracts defining the expected output and a signed lineage receipt confirming completion. Bernstein provides ready-made scenario graphs for various workflows.

Installation is straightforward via uv tool install bernstein or pipx install bernstein, followed by bernstein init and bernstein doctor to check agent authentication. The CLI can then be used to execute tasks, such as bernstein -g "fix the failing test in tests/test_foo.py".

Why it matters for builders

Bernstein provides AI builders with a robust framework for establishing governance and auditability for their AI agent deployments. The ability to define policies declaratively and have them enforced deterministically is critical for building trust and ensuring compliance, especially in complex or sensitive applications. This tool simplifies the process of creating verifiable records of agent actions, which is essential for debugging, security analysis, and regulatory adherence.

Practical impact

Developers can integrate Bernstein into their AI agent workflows to gain fine-grained control and visibility. For coding agents, this means ensuring that code generation adheres to specific standards and that all changes are logged. For other agent types, it allows for the creation of auditable evidence packs, datasets, or reports. The tool's support for air-gapped environments and local operation makes it suitable for organizations with strict security requirements. Builders can start by installing Bernstein and using the bernstein doctor command to verify their CLI agent setup, then experiment with running simple tasks to observe the generated logs and receipts.

Caveats and source limits

The project is described as being in "beta" status and is solo-maintained, indicating active development where minor versions may introduce interface changes. Developers are advised to pin their Bernstein versions for any dependencies. While the system supports over 50 CLI agents, the specific performance or compatibility nuances with each agent are not detailed in the provided excerpts. The source does not include independent benchmark results or pricing information, as it is an open-source project under the Apache-2.0 license.