agentflow-in-the-flow-optimization

【效率与协作】Decompose agent work across four specialized modules (planner, executor, verifier, generator) coordinated via evolving memory. Use Flow-GRPO to convert multi-turn sparse-reward optimization into sequential single-turn updates with outcome broadcasting, achieving 4-15% accuracy gains on benchmarks while scaling better than monolithic agent policies.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准