Every handoff I've seen fail broke in one of three ways.
Watching agents pass work to each other, the failures stopped looking random. They're the same three mistakes, over and over, and people make them too.
When you watch enough agents hand work to each other, you start seeing the same failures again and again. At first each one looks like a one-off bug. After a while you realise there are really only three, and they're the same three that go wrong when people hand work over.
1. The brief describes a world that isn't real
One lead told a pair of workers that their output belonged in a repository folder that didn't exist. It sounded completely plausible. The workers were perfectly capable. The work failed anyway, because the brief described a place that wasn't there.
This is the most common failure I see, and it's sneaky because it looks reasonable. A file name that's almost right, last month's version of a document, a folder someone assumed existed. The model fills the gap with something plausible and carries on confidently.
2. The brief leaves out what already failed
When an attempt fails and the work moves on, the output usually gets kept and the failure gets binned. So the next worker tries the same approach, hits the same wall, and you pay for the lesson twice.
The useful part of a failure is specific: what was tried, what actually happened and what that means for the next attempt. "It didn't work" teaches nothing. "On a phone, the image covered the sign-up button" saves the next worker an afternoon.
3. Old and new instructions arrive together
The third one is quieter. A worker gets a brief with several versions of the truth in it: the original instruction and the correction, last month's brand guide and this month's, four separately assembled pieces of the company's identity that don't quite agree. It follows whichever one it happens to weigh most, which is a coin flip from your point of view.
We saw this when the company's facts, voice, visual style and culture each reached workers through a separate path. Each looked fine on its own. Together they could contradict each other, and nothing recorded which version any piece of work actually used.
What isn't on the list
Notice what's missing: "the model wasn't smart enough". In nearly every handoff failure I've looked into, the receiving worker was capable of the job. It was given a wrong, incomplete or contradictory picture of the world.
That's good news, because these are structural problems with structural fixes. Better models won't solve them on their own. Better handoffs will, and they work for people too.