Beyond Code—Agentic Variation Operators for Evolving Autonomous Agent Policies

TL;DR: Could AVOs evolve not just code, but the policies and behaviors of autonomous agents in complex environments, leveraging execution feedback and domain simulators? By extending AVO to policy search, we might achieve more sample-efficient and robust policy discovery than standard RL or evolutionary policy search.

Research Question: How do agentic variation operators perform when evolving agent policies—particularly in environments where policy robustness, adaptability, and interpretability are critical?

Hypothesis: AVOs that incorporate behavioral lineage, domain knowledge, and execution feedback will evolve more robust and interpretable agent policies than classical genetic programming or RL-based policy search, especially in environments with sparse or delayed reward signals.

Experiment Plan: Implement AVO-driven policy evolution for simulated control tasks (e.g., OpenAI Gym, multi-agent Coin Game as in Kolle et al., 2024). The agent can consult past policy performance, domain knowledge, and simulation feedback to propose, critique, and repair policy changes. Compare to standard evolutionary policy search, RL (e.g., PPO), and hybrid methods (Nguyen & Luong, 2023). Metrics: performance, robustness, sample efficiency, and policy interpretability.

References:

• Chen, T., Ye, Z., Xu, B., Ye, Z., Liu, T., Hassani, A., et al. (2026). AVO: Agentic Variation Operators for Autonomous Evolutionary Search.
• Kolle, M., Schneider, K., Egger, S., Topp, F., Phan, T., Altmann, P., Nusslein, J., & Linnhoff-Popien, C. (2024). Architectural Influence on Variational Quantum Circuits in Multi-Agent Reinforcement Learning: Evolutionary Strategies for Optimization. International Conference on Agents and Artificial Intelligence.
• Nguyen, T. H., & Luong, N. H. (2023). Stable and Sample-Efficient Policy Search for Continuous Control via Hybridizing Phenotypic Evolutionary Algorithm with the Double Actors Regularized Critics. Annual Conference on Genetic and Evolutionary Computation.