CRUX Team Rebuttal: Scaffold Not the Bottleneck in AI Research Agents

random_walker · x · 2026-08-20

Responding to criticism from Sakana AI that failures in CRUX 2 were "trivially fixable" via scaffold changes, the CRUX team published a rebuttal note. They argue their scaffold already incorporated key designs from The AI Scientist (lit review, parallel exploration, logging). The observed failures in judgment, backtracking, and resource awareness stem from fundamental model limitations, not trivial engineering issues.

Original post →

More from coding & agent

coding & agent channel →