Why does an API return a normal result when an agent sends a bad tool argument?

An API rejects only what its schema can check, so a constraint that exists only in the documentation’s prose is often ignored and the call still succeeds. The SilentProbe study (arXiv:2609.00035) sent 219 schema-derived bad requests to live endpoints from 27 vendors. Where the constraint was machine-checkable, such as an enumeration, the server returned an honest error in 111 of 111 cases. Where it was stated only in prose, the call failed silently in 44 of 61. An agent then can’t tell a query that matched nothing from one the server didn’t understand: both come back as HTTP 200 with a parsable body.

The agents didn’t rescue it. Running the full loop, models noticed the silent failure in 12% of cases and repaired it in none. They told the user nothing was wrong in 41%, and invented a figure in 12%. Moving the allowed values into the schema took the failure from 88 of 88 attempts to 0 of 89.

All calls went through one aggregation layer, so this measures how vendors encode constraints, not every API. For the failure itself, see the ways a tool call can fail.