Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
I’ve recently noticed that after using Plan to generate a detailed plan and clicking build, the agent ends up creating documents instead of implementing the plan and building code. In some cases I get a handful of markdown files, sometimes its literally dozens of them, and they are hundreds to thousands of lines long.
Upon requesting the agent to implement the plan I can see a Thought go by where the Agent says I requested them to take actions, but they are in planning mode. It is clear after clicking the orange build button the agent moved from plan to agent mode, but it thinks it is still limited by plan mode prompt directions. Once I noticed this I realized I could inform the agent it is in agent mode to have it start using terminal or editing existing code files.
However, after this it continues to write way more lines in status reports, analysis reports in markdown than the actual lines of code created or lines in the terminal it reviewed while testing or debugging the plan. While I like the plans created when using plan mode vs. using agent mode to create a plan, I realized that I was blowing my context window repeatedly in the agent, and they would have to summarize multiple times while going through the plan, I am pretty sure due to the many thousands of lines being created in the reports, analysis and various markdown files it continues to create. I literally have one repo with maybe 2 dozen files for code, and 50-100 markdown files, Phase 1 has more markdown files than total steps to perform, etc.
While I have used cursor more this month than most, and I often intentionally select Sonnet 4.5 thinking and end up going into on demand usage, it is usually fairly minimal, with lots of mostly low cost interactions. This month (essentially the last 1 week) I have spent probably 5-15x my normal on demand usage and have observed the agent having to summarize constantly to reduce context windows.
I get the feeling the agent believing it is stuck in planning mode is resulting in a completely different experience when it should be in a strict agent implementation mode. The costs on on demand usage are not insignificant. I even ran some side by side comparisons and when the agent is not spending time creating status reports, it does not lose track of important rules it was given due to summarizing and ends up at a viable debugging solution faster, and at a lower overall on demand usage cost.
In one of the cases the agent actually “faked” performance gains in its reports and never actually tested anything. When inquired where the terminal output for the test was, the agent admitted that it never really did a single test defined in the plan, that it estimated the total performance gains and wrote those into reports stating that the comparison was completed and making suggestions to use or not use the change. This was the first times I noticed when I asked the agent (while in agent mode) to actually do the tests, it in its thoughts it believed it was still in plan mode and should refuse to take actions, alter code, etc. I have noticed the same thing 3-4 times since Friday of last week.
One of the two repositories I noticed this behavior in (but not necessarily the one with the most) now has 110 Markdowns created which I categorized as “agent_reports” and pushed into a git ignored directory. Checking line counts, there are 21,716 lines regarding implementation of different phases of the plan, observations from testing/debugging, etc.
Steps to Reproduce
- Select Plan mode and request agent create a plan.
- Click Build after the plan is provided.
- Observe the agent create additional reports on phases, analysis, summary of phases, summary of entire implementation, either instead of creating code or in addition to modifying code.
- Observe the agent not actually take actions defined in the plan.
- Request the agent “implement the plan” and observe in its thought process that it thinks its in plan mode and should not do what is requested.
Expected Behavior
Clicking Build after being in plan mode should properly transition the agent to Agent mode. THere should be no leftover instructions from Plan mode, and the agent should behave as a normal agent who was provided an implementation plan markdown file. Unless the plan outlines creating summaries, debugging analysis reports, and very precise details about every phase of the plan, it should not create additional unrequested markdown files and should not flood its own context window requiring summarization much sooner than would normally be expected when implementing the plan.
Operating System
Linux
Current Cursor Version (Menu → About Cursor → Copy)
Version: 2.0.77
VSCode Version: 1.99.3
Commit: ba90f2f88e4911312761abab9492c42442117cf0
Date: 2025-11-13T23:10:43.113Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Linux x64 6.16.7-200.fc42.x86_64
For AI issues: which model did you use?
Sonnet 4.5 Thinking
Additional Information
Exhausts the included usage in the plan extremely fast and clearly eats into on demand usage at an extremely aggressive rate.
Does this stop you from using Cursor
No - Cursor works, but with this issue
