yesterday i gave 1 prompt to the new feature called projectsit consumed all my tokens in no time. when i am checking the history, its showing a lot of token consumption to a very basic task
after that, it hit the limit and the work is not even done
Steps to Reproduce
Try giving a simple prompt to a project with big structure
Expected Behavior
it should have completed a simple task without firing so much multiple agents
Hey @Anmol_Arora, thanks for the report, and sorry it ate through your usage so fast - nothing failed on your end. This is how Projects works today: the main chat is a coordinator that hands your request to several agents running in parallel, and each one bills its own work, so on a big repo it adds up quickly (docs). The rows marked “Free” were a promotional credit we added when you hit the limit, so they didn’t count against your plan.
For a small or single-file task, use a regular Agent chat or a single Cloud Agent instead of a Project - much cheaper. If you do use a Project, add to your first message or the Project rules: “Use a single agent for this, and don’t start parallel or helper agents unless I ask,” and “Start every agent on the same model and effort as this Project.” You can stop agents anytime from the Project’s agent list.
A built-in spend cap is a fair ask, and we’re tracking feedback like this. Your usage limit resets with your billing cycle (shown at cursor.com/dashboard).
You told me to use General Chat for simple tasks, so why was this feature designed in a way that prevents users from working on long-term projects?
According to the documentation, this feature is supposed to support extended, ongoing work. I intend to use it for a long-term project, but naturally, I need to begin with a short and simple task. How could a basic task like this consume nearly 500 million tokens?
How are you expected to work with a feature that prevents you from working?
I gave it a general prompt to start working on a project, but the response was highly unprofessional. When starting a project, you are not supposed to consume all the available tokens at once. The project should be built gradually, step by step.
Even a simple task should not consume this many tokens, especially when no detailed context has been provided by the user. I gave a very basic instruction: find the validations in the documents and verify them against the project. My intention was to continue working on these validations over a long period, perhaps even a month.
Instead, the response focused on token usage and limitations rather than beginning the actual work. This is extremely frustrating, unprofessional, and disappointing. Knowing that this is how the feature works does not solve the problem. The system should be able to start with the available context and continue building the project gradually.
@Anmol_Arora you can use Projects for a long-term project. The issue is the opening prompt, not the feature. A broad “verify all validations against the project” on a big repo makes the coordinator start several agents at once, each spawning helpers that all read your repo into context. That repeated reading, not your instruction, is where the tokens went.
To run it gradually:
Add to the Project rules: “Use a single agent per task. Don’t start parallel or helper agents unless I ask. Start every agent on the same model and effort as this Project.”
Give one narrow task at a time (one document or module), not the whole project at once.
Stop anything unexpected from the agent list while it runs.
That drops a starting task to a fraction of the cost, and you keep building on the same Project over weeks. A built-in spend cap is fair feedback we’re tracking.
