Day-2 Operations
Keep a deployed agent good: burn-rate SLOs at a low target, detection for failures that throw no error, runbooks made of levers, and retirement as a dated migration.
Last updated
After this section you can
- Derive an error budget and burn-rate thresholds for a quality SLO
- Detect and contain a silent quality failure without a deploy
- Plan a forced model retirement as a dated migration
39
Day-2 Operations
Set SLOs that catch a failure with no error, stop it with a lever that needs no deploy, and retire a model on a dated plan.
Key idea
An agent gets worse while nothing deploys. Measure the outcome against an SLO, detect failures that throw no error, stop them with a lever, then add an eval case.
Nothing pages when an agent is wrong: a probe catches it, a lever stops it, an eval case keeps it stopped.
Pans sideways on a phone.