When you're automating a large number of workflows, whether they are based on coding or knowledge work, and constantly delegating or constructing loops with agents, the fundamental question that you need to ask is: does your intent survive through the end result?
What needs to propagate through a chain of prompts, tool calls, and lookups is your intent, as a signal that can drive downstream actions to achieve a desired outcome.
If that signal dilutes away when agents carry out work, they will inadvertently go off the rails. You need checkpoints to make sure it survives at a sufficient resolution and makes it to the next step.
Or you construct short feedback loops, ones that elicit more intent from you to produce a sharper and better outcome at the expense of less automation.
I think this is the key bottleneck to full automation when not every requirement can be specified in painstaking detail, and where inherent assumptions on both sides, human or otherwise, can snowball into something you didn't want in the first place.