Agentic Steering of Evolutionary Search for Safe-and-Sustainable-by-Design Chemicals

Chemical pollution poses a critical threat to human health and the environment. Together with the increase in chemical production this calls for a shift toward safe-and-sustainable-by-design (SSbD) methodologies. However, safety-aware generative modeling is hindered by scarce environmental endpoint data and consequently unreliable oracles for properties such as persistence and ecotoxicity, making de novo molecular optimization particularly vulnerable to reward hacking. Expert judgment can help address such underspecified objectives, but continuous human supervision is costly and difficult to scale. To address this bottleneck, we present an agentic system that augments a graph-based genetic algorithm with LLM-based strategic search steering. The LLM interprets a multi-objective design goal specified in natural language, diagnoses the evolving population, and adaptively intervenes through a set of tools that enable molecular edits, population-level steering, and revision of the numerical objective. We apply the system to the design of safe phenolic antioxidants and observe that agentic steering produces populations that avoid reward hacking patterns recurring in baseline populations, achieve a higher AiZynthFinder solve rate, and yield more structurally diverse, suitable candidates.