Longer arguments and short observations on agents, infrastructure, and what actually happens when these systems meet the real world.
ARC-AGI-3 shows the same model scoring 63% or 99% depending on what surrounds it. My own benchmark shows why that isn't the whole story.