# shepherd-agents/shepherd

A runtime substrate that turns an agent's execution into a reversible, Git-like trace, so meta-agents can observe, fork, replay, and revert any run. Couples agent and environments in a copy-on-write fork ~5x faster than docker commit, with ~95% KV-cache reuse on replay. Framework built for meta-agents to supervise, optimize, and train other agents

Repository: https://github.com/shepherd-agents/shepherd
Canonical: https://ross.abutalabs.com/products/shepherd-agents-shepherd
Homepage: https://shepherd-agents.ai/
Language: Python
License: MIT
License Family: permissive
Topics: meta-agents, meta-optimization, runtime-supervision, workflow-automation, mcts-rl, tree-rl
Last push: 2026-08-09T20:42:47+00:00

## Health v2 (maintenance only)
Score: 76/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 96, release rhythm 92, longevity 5
- inputs: {"age_days": 70, "days_push": 24, "days_rel": 56, "gap_med": 1, "n_releases_24m": 4}
- flags: young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2376, forks 205 (observed 2026-08-28T04:06:41.921366+00:00)

## What it is
Shepherd is a Python runtime substrate that records LLM agent executions as reversible, Git-like traces so meta-agents can inspect, fork, replay, and revert any run. It couples agents with copy-on-write forked environments and enforces OS-level permissions, enabling supervision, counterfactual optimization, and tree-search RL over other agents.

## Use cases
- supervise and intercept risky actions from parallel coding agents
- replay agent runs from a past state to test workflow edits
- fork agent execution to explore alternative trajectories
- train agents with tree-search reinforcement learning over execution traces
- run sandboxed agent tasks whose file changes arrive as reviewable proposals
- revert an agent's environment to any prior state quickly
- build meta-agents that coordinate, halt, or repair other agents

## When to choose
- you need runtime observability and reversibility over LLM agent executions
- you are building meta-agents that manage, optimize, or train other agents
- you want sandboxed agent tasks with OS-enforced permission grants and reviewable changesets
- you need fast forking/replay of agent environments for counterfactual or RL experiments

## When to avoid
- you need Windows support (only macOS and Linux enforce grants; use WSL)
- you need a production-stable framework (it is early alpha with changing APIs)
- you just want a simple chatbot or agent loop without trace/replay machinery

## Facets
- artifact type: framework
- maturity: experimental
- function: agent-framework, workflow-automation, mcp, tracing, llm-inference, reinforcement-learning
- domain: large-language-models, developer-tools, reinforcement-learning
- platform: python, cli
- tags: meta-agents, execution-traces, reversibility, copy-on-write, kv-cache-reuse, agent-supervision, tree-search-rl, counterfactual-optimization, permissions, sandboxed-agents, sandboxing, replay, ai-agents, automation, macos, linux

## Member repositories
- shepherd-agents/shepherd (main) score 76

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:41.921366+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:35:02.037852+00:00, confidence not recorded.
  - readme: https://github.com/shepherd-agents/shepherd (fetched 2026-08-28T04:06:41.921366+00:00, sha a181f2508327)
  - homepage: https://shepherd-agents.ai/ (fetched 2026-08-29T10:16:01.820759+00:00, sha 2e3baf2b134a)
  - site_page: https://docs.shepherd-agents.ai/ (fetched 2026-08-29T10:16:01.823345+00:00, sha 3a2b1b8dbf31)
- Data as of 2026-08-30T08:39:29.467469+00:00.