Series
Production failures
Find the failures that surface after an agent has already changed something.
5 articles
Reading list
In this series
- 01
· Operations · 6 min
Five things that keep breaking when AI agents move into production
The operator-pain pattern that keeps recurring in CISO discovery calls, with concrete scenarios and the standard fixes that don't close them.
- 02
· Operations · 5 min
Discovery lag: how long before you find out what your agents did?
Agents act at machine speed; teams find out at meeting speed. Naming the gap, measuring it, and the one mechanism that makes it negative.
- 03
· Operations · 7 min
Stopping a running agent
Three containment outcomes (stop, scope, quarantine), the credentials-in-flight problem, the sub-agent cascade, and what a kill switch actually means at runtime.
- 04
· Engineering · 6 min
A 1-second submit hid a 30-minute outage
How a read-only agent investigation separated a compound runtime outage into five causal defects, and the operating rules that came out of it.
- 05
· Engineering · 5 min
After the outage: making agent completion provable
An agent turn can be accepted, scheduled, streamed, and still fail. We changed the runtime so intermediate progress can no longer masquerade as completion.