Preprint reports agent system improved protein-folding model search under matched compute

An arXiv preprint describes AgentFold, a multi-agent system that modifies and evaluates protein-folding model code in a closed-loop search. Starting from ESMFold, the researchers say it explored about 80 variants, consuming about 5,000 GPU-hours and 170 million LLM tokens. Under a matched compute budget, it improved the best lDDT by 7.5% over independent Codex proposals and beat a random-search control. Code and experimental resources are publicly available.