Has anybody else noticed that 1.5 has regressed to a 9 year old’s level of analytical skill and contextual awareness. 2.0 is absolutely unworkable in the analysis and planning, but 1.5 has lost about 40 points on the IQ scale.
in 100% of the plans I’ve implemented over the past 3 days, the agents have either gone against my instructions and sneaked in code that does the opposite of what was instructed, or totally ignored key instructions. I even caught one trying to designate a UUID field in the user table as storage for client keys, nearly erasing all my active clients. I am actually concerned using Cursor at the moment, it is sabotaging every implementation to the point where it is having a business impact.
Hey, that sounds frustrating, especially the part about almost overwriting a UUID field. That’s a real risk.
To look into this properly, we’ll need a bit more info:
Which exact models do you mean by 1.5 and 2.0?
Are you using Agent mode, Plan mode, or both?
Can you share a Request ID from one of the sessions where this happened? top-right of chat > Copy Request ID
Do you have any project rules or a .cursorrules file set up?
Also, a concrete before and after example would help a lot. Your prompt vs what the agent actually did. Going against instructions can sometimes be about how context is being passed, especially in longer sessions where the context window gets full.
One general tip for now, if the agent is making unwanted changes, shorter and more focused chats usually work better than long-running ones where context can degrade.