Every place a program calls a model is a boundary with an agreement attached: what must hold before the call, what the request has to say, and which replies may be passed on to something that acts. This paper derived that agreement for 496 call sites in 40 open-source agent systems from the job each one does, then compared it with the code. 82% of those call sites check something before and after the call — and 5.3% cover their whole agreement. The gap is not missing effort but misplaced enforcement: two thirds of what belongs in the request is asked of the model instead of verified in code, and ten recurring shapes account for 99% of the misses.