FlowEvo: Self-Evolving AI Agents via Co-Evolution of Workflows and Skills
A new AI framework called FlowEvo enables large language model agents to automatically compile successful problem-solving traces into reusable skill records, allowing agents to improve over time without retraining. The system operates through three mechanisms: workflow-to-skill compilation, skill-to-workflow feedback, and skill bank management with safety checks. This training-free approach addresses a key limitation of current LLM agents, where useful procedures discovered during execution are typically lost after a task is completed. FlowEvo is detailed in a preprint on arXiv (2607.21596) and represents a step toward self-evolving AI systems.
Key facts
- FlowEvo is a training-free framework for LLM agents
- It compiles successful execution traces into reusable skill records
- Each skill record pairs a callable artifact with structured guidance
- Admission applies interface, replay, and safety checks
- Skills persist in a skill bank at inference time
- Three mechanisms: workflow-to-skill compilation, skill-to-workflow feedback, skill bank management
- Designed to overcome the transient nature of inference-time workflows
- Preprint published on arXiv with ID 2607.21596
Entities
Institutions
- arXiv