Large language model (LLM) agents often construct workflows at inference time to handle complex tasks, but these procedures are typically discarded after a single episode. Existing skill libraries offer reusable routines, yet they are usually built offline and fail to grow from an agent's own experiences. FlowEvo, introduced in a new paper on Hugging Face, addresses this bottleneck by enabling workflows and skills to co-evolve at inference time.
FlowEvo is a training-free framework that compiles successful workflows into callable skills, stores them in a persistent bank, and retrieves them either for direct execution or as context for constructing new workflows. It also tracks each skill's downstream utility and suppresses skills that cause negative transfer, ensuring the agent improves continuously without additional training.
Using a shared GPT-4o-mini backbone, FlowEvo achieves the highest accuracy among 8 baselines on the full standard splits of ALFWorld, HumanEval, MBPP, GSM8K, and MATH-500. On ALFWorld, it reaches 85.6% accuracy—26.4 points above the strongest baseline—while using roughly one third as many tokens. Across 10 base models ranging from 7B to 671B parameters, FlowEvo outperforms ExpeL in 49 of 50 model-dataset comparisons.
The framework's ability to turn verified workflows into persistent skills closes the loop between inference-time problem solving and long-term learning. The code is available at github.com/DEFENSE-SEU/FlowEvo.