What's the boring thing that still breaks your production agents first?

I keep a rough tally of agent incidents I hear about, and the model is almost never the cause. It's the stuff sitting next to it.

The usual suspects on my list:

- a connected system changes a field and the agent keeps going with bad data

- a retry loop quietly eats the budget

- permissions drift after a reorg

- a prompt edit ships and nobody reruns the test cases

- nobody notices for a week because the dashboard only shows uptime

I work on agent rollouts at AI Hive, so my sample skews toward regulated teams, where the audit trail matters as much as the output.

For those of you running agents live: which one bites you most, and what did you change after the first time? And what's missing from my list?