These are not 24 hours of inference with floating point errors accumulating; largely the system guards against errors compounding. Tool failures, compile failures, test failures, etc, push back against the model taking a wrong turn and force it to correct.

Yes it's much easier to have a smarter model that goes straight to the correct answer first, but it may not be necessary or economical. There's a minimum bar for the model where it understands problems and knows the right step to correct them, and above that newer models give diminishing returns.