Glossary · Term

Shepherd

← all terms

Definition

Plain language

A system that records everything an AI agent does so a second agent can inspect, copy, rewrite, or rewind it like version-controlled code.

As stated in the literature

A runtime substrate that represents an agent's execution as a formalized, append-only trace with copy-on-write forking, byte-identical replay, and effect gating, enabling meta-agents for supervision, optimization, and RL.

Why it matters: By making an agent's run inspectable and replayable like version-controlled code, it enables much stronger oversight and improvement of agent behavior.

For example, it records everything an agent does so a supervising agent can branch off a copy, rewind to an earlier point, or block a risky action before it happens.

Heard on the show

“The paper is called "Shepherd: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution Trace.”
Episode 096 — How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty

Mentioned in 1 episode

  1. 096
    How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty

Related terms