AutoRAS: Learning Robust Agentic Systems with Primitive Representations

The automated design of agentic systems offers a promising pathway for scaling large language models (LLMs) beyond single-agent reasoning. While prior work has advanced task performance through handcrafted or automatically generated multi-agent workflows, robustness is often treated as an afterthought, leaving systems vulnerable to external adversaries and internal failures. We propose AutoRAS, a framework for the Automated design of Robust Agentic Systems. AutoRAS formulates system design as generating a sequence of symbolic primitives that jointly encode structural connectivity and behavioral actions, and learns to optimize this sequence using execution-derived safety signals and flow-based sequence-level objectives. Extensive experiments show that AutoRAS achieves the best performance in both vanilla and adversarial settings, with the smallest performance degradation under attacks. Further analyses demonstrate strong transferability, stable optimization behavior, stability across primitive sets, and favorable cost trade-offs. Our code is available at this link.

When multiple agents collaborate to solve complex tasks, the resulting systems can be fragile because misleading input or faulty component can cascade through the team and produce wrong answers. Existing approaches to designing such systems often struggle to balance task performance with robustness against these failures. We introduce AutoRAS, a framework that automatically designs robust agentic systems. We represent every design choice, including how agents are connected, what roles they play, and where safety checks are placed, as a sequence of reusable building blocks. AutoRAS then searches over this space and learns from execution feedback to find designs that perform well even when parts of the system are deliberately attacked. We evaluate AutoRAS across reasoning, question answering, math, and code generation tasks, showing that our automatically designed systems consistently outperform existing alternatives in both normal and adversarial settings. Our work demonstrates that automated design can effectively produce robust agentic systems across diverse tasks.