The Agentic Idea Manager (AIM) is a fully autonomous framework for managing and exploring research directions in automated research. It distinguishes idea-driven search from solution-driven search and addresses three core challenges: organizing evolving ideas, selecting promising directions, and maintaining alignment between ideas and their implementations. Inspired by Bayesian optimization, AIM uses an Agentic Surrogate to organize discovered ideas, an Agentic Acquisition mechanism to guide selection, a Solution Auditor to check idea-solution integrity, and a Resource Planner to allocate experimental budgets adaptively.
AIM was evaluated on 10 AutoLab benchmark tasks spanning System Optimization and Model Development & CUDA. It outperformed the strongest baseline, ScientistOne, by 1.6 percentage points on System Optimization and 4.9 percentage points on Model Development & CUDA. Moreover, AIM reached ScientistOne’s best performance up to 3.1 times faster in wall-clock time. The paper also provides a theoretical analysis showing that explicit idea-level allocation makes semantic coverage controllable and becomes especially beneficial when competitive directions are sparse among plausible alternatives.