Bernstein: Open-Source Governance and Orchestration for AI Agents

Introduction

Bernstein is an innovative open-source framework that serves as the governance and orchestration layer for AI agents. It empowers users to define rules declaratively, specifying who can perform what actions, what requires approval, and what must be recorded. Bernstein then enforces these policies, generating a verifiable and replayable record of every operation. This project is free and distributed under the Apache-2.0 license, emphasizing transparency and control in AI agent deployments.

Why Use Bernstein and Key Advantages

Bernstein stands out with several core differentiators that address critical challenges in AI agent management:

• No LLM in the Coordination Loop: Unlike many other frameworks, Bernstein's scheduler is implemented in plain Python, ensuring that agent runs are reproducible end-to-end. This deterministic approach means replaying a plan yields the exact same task graph, eliminating non-determinism in coordination.
• Checkable After the Fact: Every run is meticulously recorded in a replay journal, and a continuous lineage spine tracks all lineage-bearing steps. An opt-in HMAC-chained audit log (BERNSTEIN_AUDIT=1) adds verifiable receipts that can be checked offline. This design allows for post-facto verification, identifying any non-determinism as a hash mismatch at the precise step.
• Isolated by Construction: Each coding task operates within its own isolated Git worktree, protected by merge gates. Artifact-mode tasks receive a dedicated working directory. By default, agents do not share mutable workspaces, with shared state limited to an atomically claimed task backlog. Stricter filesystem enforcement is available through sandbox backends.
• Broad and Local: Bernstein offers extensive compatibility with 52 selectable CLI agent adapters, alongside a generic --prompt wrapper. It operates with file-based state, avoiding SaaS dependencies or third-party data planes, making it suitable for air-gapped environments.

Installation

Getting started with Bernstein is straightforward. You can install it using uv or pipx:

uv tool install bernstein    # or: pipx install bernstein
bernstein init
bernstein doctor             # checks a CLI agent is installed and authenticated
bernstein -g "fix the failing test in tests/test_foo.py"

For more detailed installation instructions, including options for pip, brew, dnf, npm, Docker, and air-gapped deployments, refer to the official install guide.

Examples

Bernstein allows you to define complex agent workflows using declarative YAML files. Here is an example of a workflow designed to produce an audit evidence pack:

name: audit-evidence-pack
version: "1.0.0"

phases:
  - name: scope
    allowed_roles: [manager, architect]
  - name: collect
  - name: validate
    allowed_roles: [qa, security]
  - name: deliver
    allowed_roles: [security, manager]

nodes:
  define-control-inventory:
    phase: scope
    role: architect

  collect-audit-logs:
    phase: collect
    role: security
    depends_on: [define-control-inventory]

  # three more evidence streams collect in parallel:
  # collect-sboms-and-attestations, collect-runbooks-and-policies,
  # collect-eval-results

  assemble-pack:
    phase: validate
    role: docs
    depends_on:
      - collect-audit-logs
      - collect-sboms-and-attestations
      - collect-runbooks-and-policies
      - collect-eval-results

  mock-auditor-pass:
    phase: validate
    role: qa
    depends_on: [assemble-pack]

  remediate-findings:
    phase: collect
    role: docs
    depends_on:
      - source: mock-auditor-pass
        condition: "status == 'failed'"
    retry:
      max_attempts: 3
      until: "status == 'done'"

  sign-and-deliver:
    phase: deliver
    role: security
    depends_on:
      - source: mock-auditor-pass
        condition: "status == 'done'"

You can also verify the integrity of a run offline using a signed receipt:

bernstein verify receipt docs/assets/demo-run/run-receipt.json \
    --public-key docs/assets/demo-run/run-receipt.pub.pem

This command re-verifies the committed receipt, ensuring that the published evidence remains untampered.

Links

• GitHub Repository: https://github.com/sipyourdrink-ltd/bernstein
• Official Website: https://bernstein.run
• Documentation: https://docs.bernstein.run/