The new functionality of creating and editing agent/chat is really powerful, and so is potentially the auto-select model concept. But is there any information from the Cursor team on how the model is actually selected? What are the heuristics?
My preference would honestly just be Claude 3.5/3.7, except when there are rate limits or unavailabilities, then fallback to some other model that the user specifies. I think it should be kept simple!
Not sure everyone will agree with you to have a simple default preference that might cost double the fast requests. For simple tasks why not choose the agent-less yet 0 fast requests deekseek-v3 for Edit and Ask Modes, and the 1/3 fast requests o3-mini for Agent Mode? Then for more complex tasks, choose by claude-3.7-sonnet and as a last resort, claude-3.7-sonnet-thinking?
P.S.
Please consider renaming the title word heuristics and perhaps the entire phrase into a more SEO friendly one (e.g. “Please clarify how does Cursor auto-select a model”) so that more people will notice this thread.
what about a “Logical_Loop” whereby you spray certain things to Bot – but you can create a || PIPELINE
{promptA} → [MODEL_B] → {promptA_input} → ETC
So push a thing to model A, then pipe output to model B, etc and you have a real workFLOW as an input to a model pipes its output to the input of next model, and it logs its transactional along the way…
And you can specifiy in RULES on how to comply to the flow…
(Aside from RULES, in the UX – we need a LOGS section that is separate from RULES where LOGGING specificity is declared… (I want LOGGING to be a complete feature. Not an inline… Logging is a trace - so just blurt it out as going along, through the filter of asked.
A token tax of sorts,
Because you may reach the Fast Requests’ limit rather quickly.