EvoSkill: An Evolutionary Learning Framework for AI Agents to Auto-Evolve Skills

EvoSkill is an evolutionary learning framework that enables AI agents to auto-discover high-performance skills using evolutionary algorithms, eliminating manual prompt engineering. It treats agent configurations (system prompts + skill sets) as iterable programs, iteratively testing, improving, and retaining optimal variants. Benchmark tests show it can match or exceed manually tuned configurations, marking a shift from static design to dynamic auto-evolution.

Traditional AI agents rely on fixed prompts and predefined skills, which struggle with complex, variable tasks. Manual prompt engineering is time-consuming and fails to cover all scenarios. Different tasks demand distinct skill combinations, and human tuning often misses hidden optimization spaces—creating an urgent need for automated skill discovery.

EvoSkill's core idea: Treat agent configurations as auto-iterable programs. Its 5-stage evolution cycle:

1. Base Agent: Uses current optimal config to solve benchmarks, collecting success/failure cases.
2. Proposer: Analyzes failures to identify root causes and propose improvements (new skills or prompt changes).
3. Generator: Converts proposals into executable changes (writing skill files or revising prompts).
4. Evaluator: Tests new variants on validation sets to measure improvement (accuracy etc.).
5. Frontier Set: Maintains top N configs via Git branches, ensuring evolution toward better solutions.

EvoSkill uses a modular architecture (Python3.12+; uv/pip for dependencies) and supports multiple LLM SDKs (Claude, OpenCode) and models (Claude Sonnet, DeepSeek-V3, Gemini). It offers a simple Python API to start evolution cycles or customize parameters (iterations, frontier size, concurrency). Extensibility is enabled via API or scripts—developers can register new tasks with custom agent factories, scoring functions, and datasets.

EvoSkill was validated on benchmarks like DABStep (data analysis), SEAL-QA (search-enhanced QA), and OfficeQA. Results show auto-discovered configurations match or exceed manually tuned ones. This proves automated skill discovery's feasibility and reveals manual prompt engineering may have optimization blind spots that automated methods can explore.

EvoSkill reduces prompt engineering and skill design costs for enterprises. Real-world applications include auto-evolving customer service bots (better query handling) and data analysis assistants (more efficient processing). It represents a shift from static AI agent design to dynamic auto-evolution, enhancing long-term adaptability and value.

Limitations: High LLM call costs; limited validation on specific task types. Future directions: Optimize evolution algorithms to reduce iterations; improve evaluation strategies; explore multi-task joint evolution; integrate with MLOps workflows.