Is it possible for the agent to use the debugger in Cursor? Many times the agent tries to understand behavior of a program soley on the code - but this is actually a really hard problem especially in larger / more distributed codebases. I am pretty sure letting the agent run the program and debug its way through using tools / mcp servers and see intermediate values like we do, would allow the agent to solve many of the harder problems.
2 Likes
This would be a game changer!
I use VS Code every day and every day we go through this cycle of ālet me add some logging to understand what is happeningā and then 30 minutes later we might get an answer. If I fire up the debugger and cut and paste screen shots we get an answer in 10 minutes but that is a laborious and needless process.
If instead the agent, which already can see the code, would be able to work with the debugger or at least see it we would be done in 5 seconds or less. All of these debuggers like in VS Code are built on top of a command line debugger.
In the real world of existing code this is absolutely a necessity and would turn VS Code into a really essential part of everyday coding. AI is only as smart as the information it has and if only has a static view of the code thatās about half of what it really needs.
Give the dang AI some glasses for goodness sake.
1 Like
yes please
1 Like
The problem can be partially solved by high 75-85% test coverage. Especially if the Agent writes them himself.
I tried to setup an MCP debugger, but not luck yet. Maybe Cursor can further look ito this repo:
1 Like
Yes, if we pay the compute for infinite test coverage but certainly āautoā canāt create a simple test other than check ānot Noneā LOL. Sonnet creates good tests but for code that already exists going back and adding a bizillion tests is not practical. If I create a new module I definitely have tests but even then there are some problems that cannot be foreseen or are edge cases even in new code that could be solved in a heartbeat with AI in the debugger.
A big problem with AI today is that it assumes your writing new code when in fact most of the world is built on existing code.
I have a project Agent Enforcer but it turned out to be of no interest to anyone and has problems with the MCP part, so Iāve given up on it for now and use individual automated testing scripts for my projects.
By the way, some time after the project was published, a separate linter launch step began to be displayed in Cursor Agent Chat 
Also, I have ideas on how to evolve it into a more interesting product, but I donāt have enough time to do it 