The system resolves execution failures without continuous human oversight
Ever watch an agent solve a problem with 100% confidence and get it completely wrong?
You can fix it… if you catch it. But then you need to watch every… single… run.
New paper: We show that a meta-agent can do this for you, so your agents get ~2x more done on their own. 🧵
Users praise the Meta-Agent system for overseeing AI agents to double task completion and catch errors, calling the work very nice for its execution traces.
No Digg Deeper questions have been answered for this story yet.
Your harness should be an agent … a meta-agent
Ever watch an agent solve a problem with 100% confidence and get it completely wrong?
You can fix it… if you catch it. But then you need to watch every… single… run.
New paper: We show that a meta-agent can do this for you, so your agents get ~2x more done on their own. 🧵
Shepherd: Meta-agents can manage your agents, so you don't have to.
📄 Paper: http://arxiv.org/abs/2605.10913 🌐 Blog & more: http://shepherd-agents.ai
Team: @simon_ycl @dch @AnanjanN @dilarafsoylu @SunJiuding @chrmanning @shi_weiyan
@StanfordNLP @StanfordHAI × Northeastern
What makes this work? Our framework: Shepherd 🐑
It represents agents as Python functions & records runs in a Git-like history that other agents can work on.
And since meta-agents are also agents, you can stack them as high as you want!
💻 Quickstart: http://github.com/shepherd-agents/shepherd
📈 Agent optimizer: improves your agents from their past runs.
It proposes deltas and tests them by replaying only the affected parts of the run, with 95% KV-cache reuse.
Outperforms the current best meta-optimizer (MetaHarness) by 27% on LiveCodeBench, and is 2x as fast.
Meta-agents are agents that manage other agents. We built a new way to write them, in just a few lines of Python.
What does this unlock? 👇
🧑💼 Agent supervisor: when your agents go off track, it rewinds their runs and gets them back on track.
CooperBench: 28.8% → 54.7%
🌲 Tree search for agentic RL: a meta-agent that squeezes richer signal from each training run.
It branches the run into many continuations and sees how they end. It then uses that signal to improve credit assignment in RL training.
2x GRPO's gains on TerminalBench-2.
Many thanks to everyone whose feedback shaped this work! @lateinteraction @gneubig @MonicaSLam @ChrisGPotts @ChenSun92 @Benjamin_eecs @yi_xin_dong @PranjalAggarw16 @harshitj__ @AkshayIyer22 @ZhengxuanZenWu @ChengleiSi
And very grateful for your support @tinkerapi @modal
@shi_weiyan @simon_ycl @dch @AnanjanN @dilarafsoylu @SunJiuding @chrmanning @stanfordnlp @StanfordHAI Very nice work. To me, making execution traces a first-class object is the key idea here. Fork/replay/revert plus 95% KV-cache reuse makes supervision and optimization much more practical.
