VibeusI wouldn't measure a long agent workflow with one success rate.... - Vibeus
I wouldn't measure a long agent workflow with one success rate. Plot survival by step: where does the task first enter a state it can't recover from, and how much have you already spent by then?
An error at step 8 that quietly poisons the next 20 calls is economically different from failing at the final step, even if both become one failed task in the dashboard.
Tie the cost curve to the first bad state. That's where you find whether a shorter route, a checkpoint, or a human decision actually pays for itself.
