FlowEvo Enables AI Agents to Self-Improve by Learning Reusable Skills from Successful Workflows

FlowEvo Enables AI Agents to Self-Improve by Learning Reusable Skills from Successful Workflows

Researchers have introduced FlowEvo, a training-free framework that allows AI agents to improve over time by converting successful problem-solving workflows into reusable skills, according to a paper published on arxiv.org.

According to the research, FlowEvo addresses a key limitation in current AI agents: useful procedures discovered during task execution are typically transient and don’t systematically benefit future tasks. The framework operates through three coupled mechanisms: workflow-to-skill compilation that extracts reusable artifacts from successful traces, skill-to-workflow feedback that retrieves accumulated skills for future problem-solving, and skill curation that monitors utility and suppresses skills causing negative transfer.

In experiments on benchmarks including ALFWorld, HumanEval, and GSM8K, FlowEvo achieved “the best accuracy-cost tradeoff among the evaluated baselines,” according to the paper. On ALFWorld specifically, FlowEvo reached an 82.8% success rate—23.6 percentage points above the strongest baseline—while using less than half the average tokens per episode compared to the most efficient baseline.

The framework enables agents to “accumulate and refine task-solving capability over time without updating model parameters,” according to the researchers. The code has been made publicly available on GitHub. The work represents a shift toward allowing AI agents to build persistent, reusable capabilities from their problem-solving experiences.