FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills

Analysis

Why This Paper Matters

Large language model (LLM) agents are increasingly used for complex tasks, but they often discard useful procedures after each episode. Existing skill libraries are typically built offline and do not adapt to the agent's own experiences. FlowEvo addresses this gap by introducing a training-free framework where workflows and skills co-evolve at inference time. This is significant because it enables agents to continuously improve from their own successes without requiring expensive fine-tuning, making it practical for real-world applications where data and compute are limited.

The paper demonstrates that FlowEvo not only improves accuracy across multiple benchmarks but also reduces token usage, which is crucial for cost-effective deployment. By compiling successful workflows into reusable skills and suppressing negative transfer, FlowEvo shows that inference-time learning can be both effective and efficient. This work challenges the notion that agent improvement requires offline training, opening new avenues for lifelong learning in LLM agents.

Technical Contributions

• Co-evolution of workflows and skills: FlowEvo treats workflows and skills as mutually reinforcing components. Successful workflows are compiled into skills, and retrieved skills inform the construction of new workflows, creating a feedback loop that improves over time.
• Persistent skill bank: Skills are stored in a persistent bank, allowing reuse across episodes. This contrasts with prior methods that discard workflows after execution.
• Retrieval mechanism: Skills are retrieved either for direct execution or as context for new workflow generation, providing flexibility in how prior knowledge is applied.
• Negative transfer suppression: FlowEvo tracks each skill's downstream utility and suppresses skills that harm performance, preventing the accumulation of bad habits.
• Training-free design: The framework requires no gradient updates, making it compatible with any LLM and easy to integrate into existing agent systems.

Results

FlowEvo achieves the highest accuracy among 8 baselines on the full standard splits of ALFWorld, HumanEval, MBPP, GSM8K, and MATH-500, using a shared GPT-4o-mini backbone. On ALFWorld, it reaches 85.6%, which is 26.4 points above the strongest baseline, while using roughly one third as many tokens. This demonstrates both performance and efficiency gains.

Across 10 base models ranging from 7B to 671B parameters, FlowEvo outperforms ExpeL in 49 of 50 model-dataset comparisons, indicating strong generalization across model scales. These results suggest that FlowEvo's co-evolution mechanism is robust and beneficial regardless of the underlying model's capability.

Significance

FlowEvo represents a paradigm shift in how LLM agents can learn from experience. By enabling inference-time co-evolution of workflows and skills, it reduces the need for offline training and allows agents to adapt to new tasks on the fly. This could lead to more autonomous and self-improving AI systems in dynamic environments, such as robotics, customer support, and scientific discovery.

The token efficiency is particularly impactful for deployment, as it lowers operational costs. The framework's training-free nature also democratizes access to advanced agent capabilities, as it can be applied to any LLM without specialized hardware or fine-tuning pipelines. Future work may explore extending FlowEvo to multi-agent settings or integrating it with reinforcement learning for even greater gains.