Prof. Andreea Bobu receives AWS Agentic AI Amazon Research Award

Assistant Professor Andreea Bobu has been awarded an Amazon Research Award in Agentic AI for her proposal “Contextual Harm Mitigation and Automated Backtracking in Computer Use Agents.”

As AI systems gain the ability to operate computers autonomously, traditional safety measures that focus only on preventing harm before it occurs. Bobu’s project proposes a framework for post-execution safety, enabling agents to detect when their actions have caused harm and autonomously plan and execute remediation strategies. 

The approach combines two ideas: reflection-driven harm detection, which allows agents to reassess their actions and identify harmful outcomes, and context-aware backtracking, which enables them to safely reverse or remediate those effects. Together, these methods aim to make agentic AI systems more robust, reliable, and safe in real-world applications.

The Amazon Research Awards program provides funding and AWS credits to support academic research in areas aligned with Amazon’s mission to advance “customer-obsessed science.” The Agentic AI call for proposals supports work on scalable AI agents, human-AI collaboration, and multi-agent systems.