Language model agents that write to live databases carry a risk that read-only systems do not: an invalid input that passes the tool layer can silently corrupt data the organization depends on. This paper empirically examines where validation sits in the tool stack, using two tool-access routes that coexist in one deployed human resources agent: a direct route with minimal checking and a service-mediated route that enforces organizational rules before writing. Across 49 edge cases, the direct route silently corrupted the database in 14 cases that the service layer rejected; both routes rejected 14; both accepted the same bad input in 12, revealing a shared blind spot; and nine reflected mismatched operations between the routes. The findings suggest that data integrity depends less on model reasoning than on where and how complete validation logic is placed. A service layer helps substantially, but alone it is not a complete defense.
Keywords: Language model agents, agentic systems, data integrity, service layer design, tool use, write operations
