Bernstein 推出开源 AI 智能体治理层，主打确定性调度与可验证审计

Bernstein 项目团队发布并介绍了名为 Bernstein 的开源 AI 智能体（AI Agents）治理层。该工具的核心机制是一个确定性 Python 调度器，它能够在不依赖大模型参与协调循环的情况下并行运行多个智能体，从而确保同一计划可以逐字节复现 1。

在最新版本 v3.14.159 中，Bernstein 对系统进行了多项优化，包括改进错误提示、移除五个命令以及增加弃用窗口。此外，新版本还引入了 Muse Code 适配器、Agent Plugins v1.0.0 清单以及支持离线验证的签名运行收据 1。为了提供透明的运行证据，Bernstein 采用“策略即代码”的方式强制执行政策，并通过 HMAC 哈希链和 Ed25519 签名生成可验证的审计记录 1。

审计功能可通过 --audit 参数启用，生成的日志以 JSONL 格式写入 .sdd/audit/ 目录下，并按 UTC 日期进行分割。这些日志条目引用前一个哈希值形成链式结构，任何篡改行为都会导致验证失败；若配置了签名密钥，系统还会写入绑定日志和谱系头的 Ed25519 签名收据 1。目前，Bernstein 能够编排 Claude Code、Codex、Gemini CLI 等编码智能体，并实验性地支持 Cloudflare Workers、Durable Objects、R2 和 D1 集成 1。

Bernstein has emerged as an open-source governance layer designed to manage AI agents through a deterministic Python scheduler 1. The system orchestrates multiple agents in parallel, ensuring that identical plans can be reproduced byte-for-byte, while explicitly excluding large language models from the core coordination loop 1. This architecture allows for precise control over agent workflows without relying on non-deterministic model inference during the scheduling process 1.

To enforce policy compliance, Bernstein adopts a "policy-as-code" approach that generates verifiable audit trails 1. These records are built using HMAC hash chains and Ed25519 signatures, providing transparent evidence of execution 1. Audit logs are activated via the --audit parameter and written into JSONL files organized by UTC date within the .sdd/audit/ directory 1. Each log entry references the hash of the previous entry, creating a chain where any tampering results in verification failure 1. When signature keys are configured, the system additionally writes signed run receipts that bind the logs to lineage headers using Ed25519 cryptography 1.

The latest release, version v3.14.159, introduces several refinements including improved error messaging, the removal of five commands with an accompanying deprecation window, and new integration features 1. Notable additions include the Muse Code adapter, the Agent Plugins v1.0.0 manifest, and offline-verifiable signed run receipts 1. The platform supports experimental integrations with Cloudflare Workers, Durable Objects, R2, and D1 services 1. Additionally, Bernstein is capable of orchestrating various coding agents such as Claude Code, Codex, and Gemini CLI 1.