shepherd vs bernstein
Agent supervision and orchestration compared
shepherd and bernstein are Python frameworks for controlling and reviewing AI-agent work. shepherd focuses on reversible execution traces and changes held for review, while bernstein coordinates policy-driven workflows, including parallel tasks and auditable runs.
shepherd: Supervise Agents with Reversible Execution Traces
Shepherd records agent work as inspectable, reversible execution traces and keeps changes as proposals for review. It is aimed at developers building systems that supervise, replay, or manage the work of other agents.
bernstein: Govern and Replay AI Agent Workflows
Bernstein is a Python governance layer for AI agents, using a deterministic scheduler to run workflows, enforce policy and record verifiable results. It suits teams coordinating CLI coding agents or other agent tasks that need approvals, isolation and offline auditability.
| shepherd | bernstein | |
|---|---|---|
| Language | Python | Python |
| License | MIT | Apache-2.0 |
| Stars | 2.5k | 1.3k |
| Forks | 217 | 184 |
| Last analyzed | Oct 4, 2026 | Oct 21, 2026 |
Key differences
- shepherd records agent work as inspectable traces and keeps workspace changes as proposals; bernstein defines workflows, roles, permissions, and approval gates through declarative configuration.
- shepherd emphasizes replaying and forking individual executions; bernstein adds scheduling, task dependencies, retries, and coordination of parallel workloads.
- shepherd is aimed at developers building systems to supervise or optimize other agents; bernstein also targets teams governing engineering, research, operations, and compliance workflows.
- shepherd is MIT-licensed and reports 2.5k stars; bernstein is Apache-2.0-licensed and reports 1.4k stars.
- shepherd is in early alpha, with OS-level grant enforcement on supported macOS and Linux setups; bernstein is beta and offers opt-in stronger filesystem enforcement through sandbox backends.
- shepherd provides a deterministic provider for an offline quickstart; bernstein describes an air-gap installation profile and offline verification of signed receipts.
Choose shepherd if you…
- need to inspect, replay, or fork agent execution traces.
- want to keep generated workspace changes separate until they are reviewed and applied.
- are building meta-agents or other systems to supervise agent behavior.
Choose bernstein if you…
- need declarative workflows with roles, permissions, approvals, and completion conditions.
- want to coordinate parallel coding tasks or non-code outputs such as reports and datasets.
- need run journals, replay, and offline-verifiable receipts for review or audit.
Compare the numbers side by side, or add more projects
This comparison is generated with AI from the OSRepos analyses of both projects. Always check each project's repository and documentation before choosing.