I was AFK for a week.
Came back to make a few very basic changes.
One change was so basic that I could have used the inline edit - removing an argument from a bunch of calls in a test file.
Side note: The file is quite big - lots of tests 
I decided to go with the agent because I want to get better at using the agent instead of doing things myself or inline edit.
I tried repeatedly, first with Sonnet-4 and then with Gemini 2.5 Pro.
It appears that the agent models are now using a different model to apply the changes. I can think that makes good sense from a cost-saving perspective but it really doesn’t work.
The models fail to instruct the “apply” model specific enough, then it changes things that I did not ask for. When I point it out the agent is as surprised as me and then switches to using sed to make manual changes.
This separation of agent and apply model is not working for me.
Am I doing something wrong or is this an experiment to see if it works?
If possible, please revert.
If these are experiments, is there a way to maybe offer a slightly reduced subscription for people who want to be subjected to these experiments and then people who pay the full subscription get things that work once the beta-tester subscription has found it does not have bugs?
As things are now I am switching to doing this by hand because the experience is so disappointing.