The expectation when removing per task buffers is that some estimation errors for individual tasks will be high and others low so ultimately to an extent they'l

The expectation when removing per task buffers is that some estimation errors for individual tasks will be high and others low so ultimately to an extent they'll cancel each other out. In Goldratt's system, the risk comes from the longest (in terms of duration) chain of dependent tasks. Delays early in the chain can only be canceled out by speed ups in later dependent tasks, that's the primary motivation for that global buffer. The technique was adapted from factory production line optimization. The case can be made that it can work well for projects doing repeatable and thoroughly understood things like constructing buildings. There are reasons it's not a slam dunk for software, but IMO there's still a lot to learn from his work.