The right tool was selected but its arguments are wrong: how should validation work?

Save the actual arguments, not just the model’s explanation

Typical symptoms include numbers supplied as text, empty required IDs and dates that look plausible but violate the API format. Save an actual call instead of checking only the model’s natural-language description. After a definition changes, a client may still supply an old schema; inspect the definition actually sent to the model for this request.

Check field types and business objects separately

Check field names and required first, followed by types and enumerations. Then verify meaning: a user lookup may require a stable ID rather than a display name. Distinguish schema constraints from business rules. Date ordering and permission conditions generally still need application checks; declaring a string does not enforce every requirement.

Return precise field errors before execution

Describe the tool’s purpose and provide parameter examples, especially for confusing units and object types. Validate against the same schema before invoking the tool and return specific field errors. Explain correctable business-rule failures. Test branches with synthetic missing-field and wrong-type inputs; checking argument handling does not require creating or deleting real resources.

Verify both argument correction and the target resource

Acceptance should cover successful valid arguments, rejection of invalid arguments and correction after the model receives an error. Check that errors are not packaged as ordinary success text and that unknown fields are not silently ignored. Confirm the operation targets the intended resource: structurally valid arguments with the wrong ID are still incorrect.

Some models and services support only a subset of schema constraints; follow the actual API documentation. This article performed no real external writes and does not claim that strict structure eliminates wrong object selection. Side-effecting actions also require clear authorization and idempotency. Do not weaken validation merely to make incorrect model arguments executable.