agentflow-in-the-flow-optimization

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.

Independent GitHub snapshot · Automated classification · Original documentation governs