Our voice agent made up a delivery date and then defended it for three turns

What happens is customer asks something, agent gives a decent sounding answer, few turns later customer pushes back and the agent just doubles down bc by then its own words are sitting in context and the model treats them like they're real. One call it guessed a delivery window early on, by the end it had repeated it twice, built a whole plan around it and was apologising for delays on a date that literally never existed anywhere except its own second turn lol. and the agent sounds most confident exactly when it's just repeating itself so customers trust it more not less. Our reviewers were checking the first minute of each call and moving on so we were reviewing the only part that actually worked. Tried telling it to treat its own answers as provisional, held for a few turns then forgot like every long context instruction does. What actually worked was splitting context into what the customer said, what a tool returned and what the agent said, only the first two count as facts now. Downside is the agent contradicts itself out loud sometimes when a tool comes back with something different than what it said before, sounds bad but at least it's right. Call scores still dipped for weeks after we shipped it though. this whole thing feels wrong tbh, we fixed it by making the model trust itself less in its own conversation, i think gnani and a couple others handle this at the platform level by keeping tool state separate from conversation context which makes way more sense than what we hacked together. anyone else dealing with this kinda stuff ?