Meta-Agents Coordinate Multiple AI Agents Mid-Task Like Git
Reactions from ranked influencers
4 posts
A multi-agent supervisor: A meta-agent can coordinate multiple agents on the fly, so they don't conflict with each other. This improves pair-coding pass rate by 2x on CooperBench.
A workflow meta-optimizer: A meta-agent can also replay the agent to try out alternative paths, to optimize its workflow. This achieves faster and better meta-optimization on LiveCodeBench and Hover.
A training guide: A meta-agent can guide the RL training by providing better per-step rewards. This doubles GRPO's performance gains on TerminalBench-2.
How to build a meta-agent in just a few lines of code? Check out Shepherd: https://github.com/shepherd-agents/shepherd And stayed tuned for new features on the way!