AutoRAS: Learning Robust Agentic Systems with Primitive Representations

AutoRAS is a newly proposed framework for the automated design of robust agentic systems, focusing on optimizing sequences of symbolic primitives that encode both structural connectivity and behavioral actions. The framework leverages execution-derived safety signals and flow-based objectives, demonstrating superior performance in both standard and adversarial settings with minimal degradation under attacks. This approach is significant for practitioners as it enhances the robustness of large language models in multi-agent environments, addressing vulnerabilities that can arise from external adversaries and internal failures.