our agent kept saying "done, CRM updated" when it wasn't. how are you verifying agent actions?

setup was simple. agent reads an inbound lead, asks a couple of follow-up questions, decides if it's qualified, then calls a tool to update the CRM and assign an owner. in testing it looked fine. then we started diffing the agent's final message against what actually existed in the database, and some runs had the agent confidently saying "lead updated and assigned to sales" when the tool had returned an error. a few times the tool was never called at all, the model just decided it had enough and jumped to the summary. the fixes people usually suggest are a stricter prompt ("always call the tool"), a reflection step where the model reviews its own work, or retries. we tried the first two. both helped a little. neither made it something i'd trust unattended, because you're still asking the model to grade its own homework. what worked was kind of dull. the model doesn't get to decide whether something happened. every tool that changes state writes a log row. after the run, normal code compares what the agent claims against what the log says, and if they disagree the run is marked failed and goes to a person. the agent's summary is a draft, never the source of truth. the cost is more glue code than the demos suggest, and less of the "just let the agent figure it out" magic. how are others handling this? do you check outcomes outside the model, or has anyone got self-verification reliable enough to leave alone?