bernstein: Govern and Audit AI Agent Workflows

Overview

Bernstein is a Python governance and orchestration layer for AI agents, built around a deterministic scheduler written in plain Python rather than an LLM coordination loop. Teams declare workflow rules, roles, dependencies, and approval conditions; Bernstein schedules agent work, applies gates, and records evidence that can be checked after a run.

It targets teams that need parallel agent work without giving up control over how tasks are assigned, verified, and recorded. It supports CLI coding agents such as Claude Code, Codex, and Gemini CLI, as well as other agent workloads that produce reports, datasets, or operational evidence. Runs use local files and can be installed for air-gapped environments.

Key Features

• Declarative YAML workflows describe phases, roles, dependencies, conditional execution, and retries.
• Deterministic scheduling runs in Python, with no model deciding the coordination sequence.
• Coding tasks can run in separate Git worktrees and pass merge gates; artifact tasks can finish against a declared contract and signed lineage receipt.
• Replay journals and lineage records support verification after execution; optional audit logging adds an HMAC chain.
• Signed run receipts support offline checks, with the README noting that identity verification requires a pinned public key.
• CLI, terminal, and browser operator surfaces expose task status, and the project lists 54 selectable agent adapters.
• Air-gap installation and local file-based state are available for environments avoiding a hosted data plane.

Use Cases

• A development team can split coding work across agents while keeping tasks isolated and requiring tests or other signals before merging.
• A security or compliance team can assemble evidence from logs, policies, scans, and evaluations, then retain verifiable run records.
• A research or operations group can coordinate agent tasks whose outputs are reports, datasets, or action logs rather than code changes.
• Teams evaluating agent reliability can run tasks repeatedly and verify the resulting reliability receipt offline.

Getting Started

Install with uv tool install bernstein, then run bernstein init and bernstein doctor to check your setup. See the README and installation guide for details.

Alternatives

• agent-governance-toolkit: It governs agents through identity, sandboxing, and policy controls across frameworks, rather than Bernstein’s deterministic workflow scheduling and replay.
• intent-plane: It focuses on fail-closed authorization for irreversible actions using signed policies, rather than governing and replaying complete agent workflows.

Considerations

The README labels Bernstein beta, says it is solo-maintained, and warns that minor releases can change interfaces. Repository activity is recent and ongoing, and the measured contributor base is not concentrated around one person; issue and pull request handling has also been active. Even so, teams should pin versions and test upgrades against their workflows. Python 3.12 or newer is required; Docker is also supported. The README identifies incomplete volunteer autopilot and unavailable hosted cloud execution as limitations. Apache-2.0 applies to the code, while the project name is covered separately by its trademark policy.