How long do silent failures go unnoticed in a production AI agent?

In one documented production agent runtime, silent failures ran from 13 hours to 60 days before anyone noticed, and the length tracked the failure’s mechanism, not how complex the code was. The source is a single-author study of a personal-assistant agent that has run in production since March 2026, with 22 incidents written up as root-cause postmortems (arXiv:2606.14589).

The longest-lived failures sat in the seams between components, such as deployment topology and contracts between scripts, where no test runs. The runtime had 4,286 unit tests, and the author reports the tests, health checks, and audits stayed green through most incidents. Roughly 70% of the failures were found by a person reading the actual output and noticing it looked wrong. A retrospective audit of 15 incidents found the runtime’s governance checks prevented none of them in advance, though they blocked 87% from recurring.

This is one runtime and one author, so treat it as a case study, not a rate for agents in general. For how to look for these failures on purpose, see how to detect silent failures in production and how to measure agent reliability in production.