AI agents are increasingly consuming APIs autonomously, booking flights, processing refunds, and querying databases without a human in the loop. Current monitoring treats every correct API response as a success. But what happens between calls matters: an agent might issue a refund without first verifying identity, skip a required confirmation step, or hallucinate policy details to the customer while the API dutifully returns 200 OK.

We formalize these failures as violations of procedure contracts: typed constraints over agent-API interaction sequences. Each constraint type captures a distinct failure mode. Forbid blocks conditions that must never hold (e.g., communicating data not returned by any prior API call). Order enforces temporal sequencing (e.g., authentication must precede any write). Ground requires that agent statements to users be entailed by actual API responses. Scope restricts operations to authorized boundaries.

When we applied these contracts to frontier models on realistic service scenarios, 27–78% of “”successful”” transactions violated at least one procedural constraint. Each model fails differently, and smarter agents make subtler mistakes that are harder to detect and costlier when they reach your customers.

Attendees will learn: (1) how to derive procedure contracts for their own APIs using a two-question method (what can go wrong for any operator, and what additionally can go wrong because the operator is an LLM agent), (2) a three-step enforcement pattern for irreversible actions (execute, verify, confirm) and how grounded-response requirements make fabrication structurally impossible, and (3) why gated deployment metrics that reject corrupt successes should replace raw success rates before any agent touches production traffic.