FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
06:00 · July 27, 2026 · arXiv cs.AI RSS
Large language model agents increasingly solve complex tasks by constructing inference-time workflows that combine reasoning, tool use, and code execution. While such workflows enable flexible problem solving, the useful procedures discovered during execution are often transient: they help solve the current task but are not retained in a form that can systematically benefit future tasks. We present FlowEvo, a training-free framework that compiles successful traces into reusable skill records. Each record pairs a callable artifact with auxiliary structured guidance, and admission applies interface, replay, and safety checks where feasible. These skill records persist in a skill bank at inference time. FlowEvo is organized around three coupled mechanisms: (1)~workflow-to-skill compilation, which extracts reusable executable artifacts from successful traces; (2)~skill-to-workflow feedback, which retrieves accumulated skills to support future problem solving through either direct execution or structured context injection; and (3)~skill curation, which monitors downstream utility and suppresses skills that cause negative transfer. Through this workflow--skill--workflow feedback loop, FlowEvo enables agents to accumulate and refine task-solving capability over time without updating model parameters. Experiments on benchmarks spanning interactive environments (ALFWorld) and code/math generation (HumanEval, GSM8K) show that FlowEvo achieves the best accuracy-cost tradeoff among the evaluated baselines under our implementation settings. On ALFWorld, FlowEvo achieves an 82.8\% success rate, 23.6 percentage points above the strongest baseline, while its average token usage per episode is less than half that of the most efficient baseline. Controlled ablations confirm that each mechanism contributes to the overall result. The code is public at https://github.com/DEFENSE-SEU/FlowEvo.
Summary
FlowEvo addresses a recurring limitation in workflow-based LLM agents: successful inference-time procedures that combine reasoning steps, tool calls, and code execution tend to remain transient, forcing the agent to rediscover similar patterns on subsequent tasks. The framework converts these one-off traces into persistent skill records without any parameter updates. Each record stores a callable artifact together with structured guidance, and admission is gated by interface compatibility, replay verification, and safety checks when applicable. The resulting skills reside in a shared bank that remains available across episodes.
The system operates through three tightly coupled mechanisms. Workflow-to-skill compilation extracts executable components from successful traces. Skill-to-workflow feedback then supplies these records to new tasks, either by direct execution when preconditions match or by injecting the structured guidance as context to shape fresh workflow generation. Skill curation continuously tracks downstream performance and removes entries that produce negative transfer. Together these steps form a closed workflow–skill–workflow loop that incrementally expands the agent’s inference-time capability.
Evaluations on ALFWorld, HumanEval, and GSM8K show that the accumulated skills improve both success rate and token efficiency relative to prior interactive and workflow-optimization baselines. On ALFWorld the method reaches an 82.8 % success rate while using less than half the tokens of the most efficient comparator. Controlled ablations confirm that compilation, feedback, and curation each contribute measurably to the observed gains. The public implementation is available at https://github.com/DEFENSE-SEU/FlowEvo.
Why it matters
This research is highly relevant for Dutch AI researchers and developers working on autonomous agents, offering a cost-effective, training-free method to improve LLM workflows. Its focus on reducing token consumption aligns with the European push for sustainable and efficient AI systems.
