Tag
This paper introduces Shepherd, a functional programming model and runtime substrate for meta-agents that formalizes operations using Lean and records interactions in a Git-like execution trace. It demonstrates significant performance improvements in runtime intervention, counterfactual optimization, and RL training by enabling fast forking and replay of agent states.