FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
FlowEvo is a training-free framework that compiles successful execution traces into reusable skill records for LLM agents. It establishes a workflow-skill-workflow feedback loop to accumulate and refine task-solving capability over time without updating model parameters. The framework includes three coupled mechanisms: workflow-to-skill compilation, skill-to-workflow feedback, and skill curation. Experiments show FlowEvo achieves an 82.8% success rate on ALFWorld, 23.6 percentage points above th
Analysis
TL;DR
- FlowEvo is a training-free framework that compiles successful execution traces into reusable skill records for LLM agents.
- It establishes a workflow-skill-workflow feedback loop to accumulate and refine task-solving capability over time without updating model parameters.
- The framework includes three coupled mechanisms: workflow-to-skill compilation, skill-to-workflow feedback, and skill curation.
- Experiments show FlowEvo achieves an 82.8% success rate on ALFWorld, 23.6 percentage points above the strongest baseline, with less than half the token usage of the most efficient baseline.
Why It Matters
This research addresses a critical limitation in LLM agent development: the transient nature of workflows discovered during execution. By enabling agents to retain and systematically reuse successful procedures, FlowEvo offers a path toward more efficient and capable autonomous systems that improve over time without costly retraining. The demonstrated improvements in both success rate and computational efficiency make this approach highly relevant for deploying LLM agents in real-world applications.
Technical Details
- Workflow-to-skill compilation: Extracts reusable executable artifacts from successful traces, pairing them with auxiliary structured guidance. Admission applies interface, replay, and safety checks where feasible.
- Skill-to-workflow feedback: Retrieves accumulated skills from the skill bank to support future problem solving through either direct execution or structured context injection.
- Skill curation: Monitors downstream utility and suppresses skills that cause negative transfer, ensuring only beneficial skills persist in the skill bank.
- Training-free operation: The entire framework operates at inference time without updating model parameters, making it computationally efficient and easy to deploy.
- Benchmark performance: Achieved 82.8% success rate on ALFWorld (23.6 percentage points above strongest baseline) with less than half the token usage of the most efficient baseline. Also tested on HumanEval and GSM8K for code/math generation tasks.
Industry Insight
- Cost-effective agent development: FlowEvo demonstrates that significant improvements in agent capability can be achieved without expensive retraining, making it an attractive approach for organizations looking to deploy LLM agents efficiently.
- Long-term agent learning: The skill bank approach enables agents to accumulate knowledge over time, creating a path toward more autonomous and capable systems that improve with experience rather than requiring constant retraining.
- Safety and reliability: The inclusion of safety checks and negative transfer suppression in the skill curation mechanism addresses important concerns about deploying autonomous agents in real-world environments, potentially accelerating adoption in regulated industries.
Disclaimer: The above content is generated by AI and is for reference only.