Plans are a best guess made with incomplete information
Go back to the kitchen renovation plan from lesson two: remove cabinets, plumbing, install cabinets, countertops, paint. That plan was reasonable on day one. But suppose the plumber discovers, mid-renovation, that a pipe behind the wall is corroded and needs replacing first. Nobody wrote that step down at the start, because nobody knew about it at the start. If you rigidly followed the original plan anyway, installing cabinets on schedule right over a pipe that needs replacing, you'd cause real damage. The plan wasn't wrong when it was made, it was made with the information available at the time, and that information changed.
This is true of essentially any plan that spans real time: new information arrives, an assumption the plan depended on turns out to be false, or an earlier step's actual result changes what the next step should even be. A long-horizon agent that treats its original plan as fixed and unquestionable, no matter what it learns along the way, will confidently keep executing a plan that no longer fits reality. That's not caution, it's just a slower way of being wrong.