Brilliant effort worth checking out. Ultra-long horizon coding tasks are where frontier models like Fable 5.1 will shine. But that's a crazy gap (over ~25 percentage points). What I think could be interesting is seeing results for a mixture of agents like what Cursor did.