Eigen RadarAI
Archive
Analysis

AgentFold searched 80 protein-model variants as Co-Scientist ran closed-loop lab work

AgentFold, on arXiv 29 August, raised best lDDT by 7.5 per cent over Codex proposals after exploring about 80 variants on a 2,000-line codebase from ESMFold. THE DECODER on 28 August reported Google DeepMind's Co-Scientist now plans experiments, runs lab equipment and drafts papers. Both turn agents into closed-loop scientific search.

Artificial Intelligence··Night
A compact lab gantry lifts one turquoise protein ribbon model into an inspection light above irregular clusters of colorful folded models on a bright workbench.

AgentFold explores 80 folding-model variants under matched compute

AgentFold, published on arXiv on 29 August, turns folding-model development into a closed-loop search over executable code variants: it proposes hypotheses, edits and debugs the code, evaluates variants, and keeps failed as well as successful interventions in a structured memory. Starting from ESMFold, the system explores about 80 variants on a codebase longer than 2,000 lines. Under a matched compute budget it raises the best lDDT by 7.5 per cent over independent Codex proposals. The search spends about 5,000 GPU-hours and 170 million model tokens, with an MCTS-style policy allocating resources. It is also measured against a random-search control and comes out ahead of that. The work is a preprint that has not been peer reviewed, and the gain was measured on one codebase and one starting model.[1]

Co-Scientist plans experiments and drives lab equipment

Google DeepMind has extended Co-Scientist, introduced in February 2025 on Gemini 2.0 as a hypothesis generator, into a closed-loop research system: it derives hypotheses, builds experimental plans, writes code, drives lab equipment, analyses results and drafts manuscripts. THE DECODER reported on 28 August that in materials science it designed synthesis recipes for two-dimensional materials and obtained semiconductor thin films on the first attempt.[2]

Code search and lab loops share a closed-loop agent pattern

AgentFold iterates executable code variants with structured memory of failed and successful edits on one protein-folding codebase. Co-Scientist closes the loop from hypothesis through lab hardware to manuscript drafts in materials science. AgentFold reports a 7.5 per cent lDDT gain on one preprint codebase; Co-Scientist's film synthesis success is reported by THE DECODER without independent replication in the pool.[1], [2]

References

  1. News sourcearXivAn agent searched 80 variants of a protein-folding model and lifted lDDT by 7.5 per cent↩1↩2
  2. News sourceTHE DECODERCo-Scientist now plans the experiment, runs the lab equipment and writes the paper↩1↩2
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →