Shepherd: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution Trace
Researchers have introduced Shepherd, a novel functional programming model designed to formalize meta-agent operations on target agents as functions, with core operations mechanized in Lean. This system records every agent-environment interaction as a typed event within a Git-like execution trace, allowing any past state to be forked and replayed efficiently. Shepherd demonstrates significant performance improvements, forking agent processes and filesystems five times faster than Docker while achieving over 95% prompt-cache reuse during replays. The model's efficacy is validated through three key applications: runtime intervention, where a live supervisor increased pair coding pass rates from 28.8% to 54.7% on CooperBench; counterfactual meta-optimization, which outperformed baselines by up to 11 points across four benchmarks while reducing wall-clock time by 58%; and Tree-RL training, improving TerminalBench-2 performance from 34.2% to 39.4%. These results position Shepherd as an efficient infrastructure for programming meta-agents. The team, including authors from prominent academic institutions, has open-sourced the system to facilitate future research in artificial intelligence and agent-based systems.
Wire timeline
Shepherd: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution Trace
Researchers have introduced Shepherd, a novel functional programming model designed to formalize meta-agent operations on target agents as functions, with core operations mechanized in Lean. This system records every agent-environment interaction as a typed event within a Git-like execution trace, allowing any past state to be forked and replayed efficiently. Shepherd demonstrates significant performance improvements, forking agent processes and filesystems five times faster than Docker while achieving over 95% prompt-cache reuse during replays. The model's efficacy is validated through three key applications: runtime intervention, where a live supervisor increased pair coding pass rates from 28.8% to 54.7% on CooperBench; counterfactual meta-optimization, which outperformed baselines by up to 11 points across four benchmarks while reducing wall-clock time by 58%; and Tree-RL training, improving TerminalBench-2 performance from 34.2% to 39.4%. These results position Shepherd as an efficient infrastructure for programming meta-agents. The team, including authors from prominent academic institutions, has open-sourced the system to facilitate future research in artificial intelligence and agent-based systems.
cs.AI updates on arXiv.org