The problem with this approach, is the probabilistic nature of the models, which results in non-deterministic behavior. They CAN directly reference the url….the question is: WILL THEY? That is the problem I have had, with the general loss of @ commands in the prompt editor. The @ commands are explicit and deterministic. We have them for files, we’ve lost them for most other things. One of the biggest losses was @code, which I used a ton in the past to explicitly target the EXACT code I needed the agent to address. I get non-deterministic results now, between the autocomplete being EXCESSIVELY inconsistent, barely recognizing a tenth or so of my codebase, and seeming to be highly biased to recent context, not to mention a significantly higher token burn as the agent “figures out” what “it thinks you mean”….
I don’t like all the non-deterministic nature that Cursor 2.x has brought to the table. Not everyone who uses Cursor is a vibe coder. Some of us are seasoned software engineers using agents to increase our productivity, and we are concerned about accuracy, precision, consistency and cost. The current Cursor 2.x approach is non-deterministic, inaccurate, imprecise, inconsistent and costs a lot more.
Please bring back the @ tools in the prompt. They give us developers who are not vibe coding, and who have concerns about cost, the ability to be explicit and precise about our intent, so that the model does not have to waste time trying to figure out what our intent MIGHT be, in their highly non-deterministic manner. OF all the tools that have cost me severely since they were lost: @code and @docs are by far, without question, the two most valuable tools that had SIGNIFICANT positive impact on the accuracy, precision and consistency of Cursor’s agent and the quality of the work it produced. The are essential, and cannot be replaced by arbitrary agent reactions to pure text prompt.
An @web tool would ensure, 100% of the time, that the agent does what we are asking it to do, rather than deliver inconsistent behavior with regards to how to handle a url. FWIW, not every url pasted into the prompt editor, is intended to be read. Some are in fact config values, or perhaps intended as input to another process you are asking the agent to execute. A tool like @web also eliminates ambiguity!
Cursor Team, please listen to and understand the needs of your users here. Not all of us are raw vibe coders. Some of us need deterministic behavior from this tool we use every single working day of our lives, and Cursor 2.x has cost us a lot of that determinism. We need it back! Please hear us!
Thanks for the report!
