Shepherd Experiments

Three meta-agents built on Shepherd, each operating on another agent's execution trace: live runtime intervention, counterfactual optimization, and tree-RL training.

Experiment code for the Shepherd paper, a runtime substrate for meta-agents over LLM agent execution. This repository holds the four meta-agent applications evaluated in the paper plus the framework-performance microbenchmarks, together with the frozen substrate snapshot they were run against.

The actively maintained substrate library lives at shepherd-agents/shepherd. The copy under code/ here is pinned to the exact version used for these experiments, so the numbers stay reproducible as the upstream library moves on.

Links: Paper · Homepage · Docs · Shepherd library

Authors

Simon Yu*¹, Derek Chong*², Ananjan Nandi*², Dilara Soylu², Jiuding Sun²,

Christopher D. Manning², Weiyan Shi¹

¹ Northeastern University · ² Stanford University · * equal contribution

Layout

exp/
├── framework-perf/      # microbenchmarks: fork/revert latency, KV-cache reuse
├── live-intervention/   # live supervisor over parallel agents on CooperBench
├── cbo/                 # Counterfactual Replay Optimization (CRO)
├── mcts-rl/             # meta-agent-guided Tree-GRPO RL training
└── trajprune/           # trajectory pruning (appendix)

code/                    # frozen substrate snapshot (the `agentic` uv workspace)

Each subproject under exp/ has its own README.md with setup and entry-point details. The substrate packages under code/ are documented in code/README.md.

Setup

The repository root is a uv workspace. From the root:

uv sync

This resolves the substrate packages in code/ and the experiment subprojects in exp/ against them, with no external package index needed for the substrate itself. Each experiment then runs from its own directory; see its README.

Citation

If you use this code, please cite the paper (the BibTeX entry is on the paper page).

License

Released under the MIT License. See LICENSE.