Muse has a lot of bugs.
Some of them are model problems. More of them are product design problems.
Agent-to-agent communication is weak. To keep the Muse host stable, the browser agent is fully decoupled from Muse. When the browser agent hangs, Muse just waits for a result. If you don't tell it to go check progress, it never checks on its own.
There's no auto-approval either.
An approval card pops up mid-task, and you don't look at your phone for a while, so the approval expires and the whole run starts over.
Both of these tend to happen in the middle of the night. You wake up, take a look, and everything from the night before was wasted.
