New Stanford, Yale and other top Chinese univ paper treats scientific model development as search over executable code snapshots, with every successful or failed intervention becoming evidence for the next branch.

Its a closed-loop agent that beat independent Codex proposals on ESMFold search.

Starting from a compact ESMFold-derived model, agents propose architecture changes, implement and debug them, run training, analyze outcomes, remember both wins and failures, then use tree search to decide which branches deserve more compute.

– arxiv. org/abs/2608.26747

Title: "AgentFold: Closed-Loop Agentic Search for Protein Folding Model Design"