It would be incredibly useful if Cursor AI allowed users to specify different AI models for “Plan” and “Act” modes—similar to Clinebot’s approach (see this discussion). This would enable more efficient and tailored workflows, optimizing for both quality and cost.
Background & Motivation
Advanced agent workflows often split tasks into two phases:
Planning:
The agent analyzes, strategizes, or drafts a step-by-step approach.
Acting:
The agent executes the plan—generating code, making changes, or otherwise “doing the thing.”
Clinebot’s feature allows you to use a high-quality model (like GPT-4) for “Plan,” and a more cost-effective model (like GPT-3.5-turbo or a local model) for “Act.” Cursor AI could benefit greatly from a similar capability.
Proposed Solution
Configurable Model Selection:
Allow users to choose different AI models for each mode: “Plan” and “Act”.
Flexible Configuration:
This could be done via the UI or a configuration file. For example:
model:
plan: gpt-4
act: gpt-3.5-turbo
Graceful Fallback:
If unspecified, default to using the same model for both modes, preserving traditional behavior.
Benefits
Cost Efficiency:
Save on API costs by using heavier models only where they matter most.
Performance Tuning:
Optimize planning for intelligence, action for speed.
Custom Workflows:
Enable power users to mix-and-match models (cloud, open-source, local) for maximum flexibility.
This feature channels a classic engineering truth: always use the right tool for the job. Cursor AI’s modular nature seems well-suited for this upgrade, which would benefit both tradition-respecting engineers and those chasing the bleeding edge.
great idea, HOWEVER, important part is: dont just make two chat modes with different permissions, make them useful. so the “act” mode should not only implement these changes but also test them and if they are not up to the requirements, it will loop until it makes them perfect. otherwise it is useless.
@ajaykumarmandapati thank you for your feature request and welcome to Cursor Forum.
Technically this is already available with Custom Modes were you can set one for Plan and another for Act with different models preselected, it also supports separate custom instructions in each mode. Custom Modes may need to be activated in Cursor settings.
I agree also with @liquefy that it would benefit from optimizing the AI actions for each mode. Upvoted the feature request!
Cursor Team will review the request and consider it.
That’s great news! Is there some documentation or how-to for this approach. Afaik there is no native support for plan and act mode but it’s achieved via Cursor rules.
ahh you mean this bit where you could tailor a custom model ( agent mode ) but how would you tie it to agent vs plan model which I’ve explicitly written in an cursor rule.
I could share the cursor rule if you wish. Afaik these custom models are specific to the current agent ( chat window ) they cannot dynamically switch between models in between a current task/chat.
Create a Plan custom mode, add most important process details to its instructions.
Create an Act custom mode, add most important process details to its instructions.
In either .cursorrules, User or Project rules you can add additional overview instructions of Plan vs. Act behavior but do not inform agent it can switch with a keyword ‘plan’ or ‘act’.
Agent can not autonomously switch between Plan vs Act in this way. You would use the custom Plan mode to start planning, Agent should be instructed to create md files based on plan. Once you are satisfied with planning you can start new chat and switch to custom Act mode, tell Agent to implement the plan and mention the md files.
Usually its best to separate plan docs in folders by feature. This lets AI stay more focused on task and reduces potential context needed to understand the task.
Also manual switching and new chat clears context e.g. while planning you may discuss partially contradicting info until you settle on a final version. In a new chat the Act mode wont have to deal then with the contradicting parts of previous Plan session. For smaller features you can manually switch the plan/act mode within a chat.
Its sure possible also to use one rule mdc file for plan vs. act but I have sometimes caught AI pretending to have received user prompt to switch mode. Usually a hallucination after longer chat sessions or when it interprets parts of requests as instructions to act.
It does depend on project, programming language and framework.
For example, I used Plan vs Act in the past intensively and had very detailed rule files. This was necessary for Sonnet 3.5 as it needed a lot of guidance.
Sonnet 4 in comparison does not work well with so detailed rules as it has a lot more of best practices & programming knowledge in its training data and reinforcement learning. Therefore for Sonnet 4 I removed most of my rules and have very basic instructions. (I mention programming language version, framework, follow SOLID & DRY, not to use interactive CLI commands)
here is the documentation and have a look then in Cursor on Agent right-click to create custom modes once activated.