收录解读
AFlow tackles a practical bottleneck in agent systems: building strong workflows still takes substantial human effort, and existing automatic approaches often rely on manually seeded structures. The paper turns workflow optimization into an explicit search problem over code-represented workflows.
Its central mechanism is to use Monte Carlo Tree Search over workflow code, with iterative modification, execution feedback, and a structured experience tree. This gives the system a concrete way to explore agent pipelines as programs instead of relying on one-shot prompt rewrites or static templates.
For the repository, AFlow is worth collecting because it is one of the clearest strong method papers in the workflow-optimization line. It operationalizes the broader ADAS / optimizable-graphs intuition into a concrete search algorithm that can improve real workflows across multiple benchmarks.
It is not ranked higher because it is narrower than the strongest route-defining abstractions and remains one method inside the broader automated-agent-design space. But as a practical and reusable workflow-search pattern, it is clearly above the inclusion bar.