LLM Agents Forge Feasible Molecules: How Route-Native Evolution Exposes Limits of Standalone Generative Models

The arXiv preprint 'My Chemical Harness' (v1, 8 Jun 2026) presents an evolutionary framework where LLM agents act solely as high-level strategy controllers over populations of executable synthetic pathways, rather than generating molecules directly. Routes are constructed from purchasable building blocks via deterministic reaction templates, validated by chemistry tools, and scored against oracles including sEH activity, synthetic accessibility, and AiZynthFinder success. On this soluble epoxide hydrolase proxy task, the hybrid agent outperformed both single-pass LLM prompting and purely deterministic controllers. Methodology relies on a modest experimental setup with fixed population sizes and oracle evaluations; no wet-lab synthesis or large-scale prospective validation is reported, and the work remains an unreviewed preprint. This approach addresses a recurring failure mode seen in earlier generative models such as those in Sanchez-Lengeling et al. (2018, ACS Central Science) and the ChemCrow framework (Bran et al., 2023, arXiv:2306.06804), where plausible-looking structures often lacked feasible routes. By constraining the LLM to preference selection over route length and reaction families while delegating execution to code, the system reduces hallucinated chemistry yet inherits the narrowness of its proxy metric and building-block library. The result suggests constrained agents can accelerate discovery pipelines, but scaling to novel therapeutic targets will require integration with active learning loops and experimental feedback absent from the current study.