Benchmarking and behavioral characterization of LLM agents for protein design

Large language models (LLMs) are increasingly deployed as agents for scientific discovery, but standardized frame-works for evaluating their performance and behaviour in scientific workflows are lacking. Protein design provides a demanding test case because modern workflows combine stochastic generative models, structure prediction systems, and physics-based evaluation tools that require extensive candidate exploration and filtering. Here we introduce BioDesignBench, a benchmark of 76 expert-curated protein design tasks spanning antibodies, enzymes, fluorescent proteins, binders, and scaffolds, together with human and non-LLM baselines and behavioural metrics derived from tool-use traces. We evaluate four frontier LLM agents across diverse protein design workflows and find that the strongest agents surpass deterministic hardcoded pipelines but consistently underperform expert practice. Although agents generally select appropriate tools, they evaluate candidate designs too shallowly, rarely compare alternatives, and terminate exploration prematurely. Guided workflows improve tool coverage but not evaluation depth. Enforcing deeper multi-metric evaluation substantially improves agent performance, demonstrating that these limitations are behavioural rather than fundamental capability constraints. We release BioDesignBench, open-source reference agents, and a public leaderboard as a community resource for evaluating and improving AI agents for protein engineering.