It is not just plan mode, FWIW. I have noticed that todo lists seem broken across the board. Not only to tasks not get completed when the agent mode is using a todo list, but it also seems like the todo list is often shared across chats, and it seems to grow perpetually, accumulating more and more todo items over time, many of which never get marked as done (even when they do get completed by the agent.)
Yeah, exactly — I’ve seen that too. You pretty much have to keep reminding the model to go back and review whether it actually completed the task correctly or all the way through. Sometimes it’ll half-finish something, skip steps, or even create a new to-do list before finishing the existing one.
It feels like the task-tracking logic isn’t fully persistent, so unless you explicitly tell it to audit or confirm completion, it just moves on and keeps stacking items.
The last few times I’ve used plan, the agent creates an .md document, but doesn’t populate it with any data. I ask why the plan is empty and they respond with bewilderment and try to add data to the md file. After that fails, they print it as a response that I then have to copy and paste into the md file to have access to the build option.
In my experience it is not rare for the model to just ignore its todo list (in agent mode) and either leave it unfinished, despite tasks being done, or just completing all the todo list tasks at once right before the chat ends.
It’s particularly annoying when the agents runs out of tokens. I want to start a new agent with the plan.md file but how does it know where to start from?!
btw: Wouldn’t it make sense for cursor/agent to start a new agent (or summarize previous work) for each todo. I was really surprised to see an agent run out of tokens (without chatting to me)
