AutoRAS: Learning Robust Agentic Systems with Primitive Representations

Researchers introduce AutoRAS, a framework for automatically designing robust multi-agent AI systems that maintain performance under adversarial attacks. The approach uses symbolic primitives to encode agent structure and behavior, optimizing for both task success and system resilience rather than treating robustness as an afterthought.

AutoRAS addresses a critical gap in multi-agent LLM systems: robustness. While prior research has focused on scaling performance through automated workflow generation, security vulnerabilities and failure modes have received limited attention. This framework treats system design as an optimization problem, generating sequences of symbolic primitives that define both how agents connect and what they do, then refining these sequences using safety signals derived from actual execution.

The research emerges from growing recognition that LLM-based agents operating in complex, multi-step workflows need defensive mechanisms. As organizations deploy agentic systems in production environments, adversarial attacks and cascading failures pose real operational risks. AutoRAS's flow-based optimization approach allows the system to learn configurations that gracefully degrade under attack rather than catastrophically failing.

For developers and enterprises building AI agent infrastructure, this work provides both practical methodology and validation that robustness-first design is achievable without sacrificing performance. The demonstrated transferability across primitive sets and favorable cost trade-offs suggest the approach scales to real-world deployment scenarios. The benchmarking against both vanilla and adversarial settings establishes meaningful performance baselines for evaluating future agentic systems.

The broader implications extend to AI safety discourse. AutoRAS demonstrates that robustness can be incorporated during system design rather than retrofitted afterward, potentially influencing how organizations architect their AI infrastructure. As multi-agent systems become more prevalent in critical applications, this research direction could shape industry standards for building trustworthy autonomous systems.

• →AutoRAS optimizes multi-agent systems for both performance and adversarial robustness using symbolic primitives and execution-derived safety signals.
• →The framework shows minimal performance degradation under attacks compared to baseline systems, indicating practical resilience gains.
• →Learned agent configurations transfer effectively across different primitive sets, suggesting the approach generalizes beyond specific implementations.
• →System design formulated as sequence optimization enables cost-efficient discovery of robust agentic architectures.
• →Research demonstrates robustness can be achieved during initial design rather than added as a post-hoc security layer.