Multi-Model Orchestration

Feature request for product/service

Cursor IDE

Describe the request

Feature Idea: Intelligent Multi-Model Orchestration Throughout the Development Lifecycle

Instead of requiring developers to choose a single model for an entire coding session, the IDE could introduce a Task-Oriented Model Orchestrator that automatically assigns the most suitable model to each stage of the development workflow.

Different models excel at different tasks. Some are optimized for long-context reasoning, others for fast iteration, code generation, or reviewing changes. Rather than treating model selection as a manual decision, the IDE should dynamically switch between models while preserving the same conversation, project context, and memory.

A typical development workflow could look like this:

1. Planning & Requirement Analysis

Preferred Models: Strong reasoning models (e.g., GPT-5, Claude Opus, Gemini Pro)

Responsibilities:

  • Understand feature requirements.
  • Break large problems into smaller implementation tasks.
  • Design application architecture.
  • Identify dependencies and edge cases.
  • Produce implementation plans and milestones.
  • Estimate implementation complexity and potential risks.

The output from this stage becomes structured context for subsequent implementation tasks.


2. Repository Understanding & Context Collection

Preferred Models: Fast long-context models

Responsibilities:

  • Index the repository.
  • Locate relevant files.
  • Understand module relationships.
  • Trace function calls and dependencies.
  • Retrieve only the necessary project context.

This stage minimizes unnecessary token usage by collecting only the files and symbols required for the current task.


3. Code Implementation

Preferred Models: Code-specialized models

Responsibilities:

  • Generate production-ready code.
  • Follow project conventions.
  • Modify multiple files consistently.
  • Create tests alongside implementation.
  • Handle repetitive boilerplate efficiently.

The orchestrator could divide larger implementation tasks into smaller units while maintaining shared context.


4. Debugging & Root Cause Analysis

Preferred Models: Deep reasoning models

Responsibilities:

  • Analyze stack traces.
  • Interpret runtime errors.
  • Trace execution paths.
  • Identify root causes instead of symptoms.
  • Suggest minimal, targeted fixes.
  • Explain why failures occur.

Rather than simply proposing changes, the model focuses on understanding the underlying issue before making modifications.


5. Code Review & Quality Assurance

Preferred Models: Models optimized for analysis and critique

Responsibilities:

  • Detect potential bugs.
  • Identify security vulnerabilities.
  • Suggest performance improvements.
  • Check maintainability and readability.
  • Flag architectural inconsistencies.
  • Validate adherence to coding standards.

This stage acts as an automated peer review before code is finalized.


6. Refactoring & Optimization

Preferred Models: Strong reasoning + code generation models

Responsibilities:

  • Simplify complex implementations.
  • Remove duplicate logic.
  • Improve performance.
  • Modernize outdated code.
  • Recommend better design patterns.
  • Preserve functionality while improving code quality.

7. Documentation Generation

Preferred Models: Fast language models

Responsibilities:

  • Generate API documentation.
  • Write inline comments.
  • Create README updates.
  • Produce migration guides.
  • Summarize implementation decisions.
  • Generate changelogs.

Using a lightweight model here reduces costs while maintaining high-quality documentation.


8. Test Generation & Validation

Preferred Models: Code-focused models

Responsibilities:

  • Generate unit tests.
  • Create integration tests.
  • Suggest edge cases.
  • Improve existing test coverage.
  • Identify untested code paths.

9. Final Verification

Preferred Models: High-reasoning models

Responsibilities:

  • Verify that all requirements have been satisfied.
  • Compare implementation against the original plan.
  • Detect regressions.
  • Ensure consistency across modified files.
  • Produce a final implementation summary.

Intelligent Model Routing

Instead of relying on fixed mappings, the orchestrator should evaluate every request using factors such as:

  • Task type.
  • Codebase size.
  • Required context length.
  • Reasoning complexity.
  • Latency requirements.
  • Cost constraints.
  • Remaining token budget.
  • User preferences (quality-first, speed-first, or cost-first).

It can then automatically select the most appropriate model—or even chain multiple models together for a single task.

For example:

Feature Request

  • Planning → GPT-5
  • Context Collection → Gemini
  • Implementation → Claude Code
  • Review → GPT-5
  • Documentation → Gemini Flash

Bug Fix

  • Log Analysis → GPT-5
  • Root Cause Investigation → Claude
  • Code Fix → Claude Code
  • Validation → GPT-5
  • Test Generation → Gemini

Large Refactor

  • Architecture Analysis → GPT-5
  • Repository Mapping → Gemini
  • Incremental Refactoring → Claude Code
  • Performance Review → GPT-5
  • Documentation → Gemini Flash

The orchestration should be completely transparent to the developer. From the user’s perspective, there is a single conversation and a single AI assistant. Behind the scenes, however, the IDE intelligently routes each stage of the workflow to the model best equipped for that task, balancing quality, speed, cost, and token efficiency.

This transforms model selection from a manual configuration into an automated optimization layer, allowing developers to focus on building software rather than deciding which model to use at every step.

TLDR: Multi agent system instead of one prompt one model breaking/segmenting task and triggering dedicated model for specific task.

Hey, thanks for the detailed feature request. Most of what you’re describing already works via Auto and Cursor Router.

Basic Auto is available to everyone, including Free. It automatically picks the right model for your task inside a single conversation while keeping context. From the user’s point of view it’s one chat and one assistant, and the model choice happens under the hood. That’s the transparent optimization layer you described.

The more advanced option is Cursor Router with preference modes Intelligence, Balance, or Cost, which matches your quality-first, speed-first, cost-first idea. One important note is that Router is currently only available on Teams and Enterprise plans. On Teams it’s enabled by default, and on Enterprise an admin enables it in the dashboard. These modes aren’t available on individual plans yet. More details: Introducing Cursor Router · Cursor

A couple options for more explicit control by role:

  • Custom subagents: you can pin a specific model via model: frontmatter in .cursor/agents/*.md like a dedicated agent for review or docs. Just note the model must be available on your plan. Docs: Subagents | Cursor Docs
  • Plan mode: separates planning from execution, so you can work out the plan first and then run implementation. Plan Mode | Cursor Docs

What you don’t have yet from your description is a deterministic pipeline that strictly chains multiple models by stage plan → implement → review → docs using your own stage-to-model rules. Similar requests are discussed here: Personalized Auto Mode and here: # Feature Request: User-configurable model routing rules for the main agent

hey dean,

thanks for detailed explanation on the existing features and sharing link to dedicated articles for the same!

it will be great if we can see cursor router available for other plans?,
moreover setting up deterministic pipeline would help in better optimizations of token in a well distributed manner

most of the model usecases by what I read on forum:

  • users don’t know which model fits in the best with best setting (context, effort)
  • if we have clearer pipelines for other plans it would be fun to use multiple model pipeline
  • in a more standard way breaking the task → recommending dedicated model for the task → choosing/setting them up and seeing the result

Glad the previous breakdown helped.

On your points:

Router on Individual plans is a reasonable request, I passed it to the team as feedback. Right now it’s only available on Teams and Enterprise, and there’s no exact timeline yet for expanding it to Individual. The easiest way to track updates is the main Router thread: Introducing Cursor Router

A deterministic pipeline a strict plan → implement → review → docs flow with your own stage-to-model rules is exactly what we don’t have out of the box yet. Similar requests are being tracked here: # Feature Request: User-configurable model routing rules for the main agent I’d recommend taking a look and adding your vote and details so the signal is clearer.

Also, your note that users don’t always know which model and which settings context and effort fit a task best is useful feedback, I passed that along too.

Until we have a deterministic pipeline, the closest practical options are what I mentioned before: custom subagents with model: frontmatter for specific roles like review and docs, plus Plan mode to separate planning from execution. Let me know if you’ve got questions about any of that.

Really appreciate you passing the feedback along to the team! Thanks for taking my ideas into consideration would love to see them integrated into Cursor sometime soon.

I think the bigger opportunity here is making this system more native and opinionated where Cursor can understand the task, break it into stages, recommend the right model/settings for each stage, and execute the pipeline deterministically.

Hey, glad the breakdown was helpful. Your vision for a native and opinionated version, where Cursor understands the request, splits it into stages, recommends the right model and settings for each stage, and runs the pipeline deterministically, is exactly the direction I shared with the team as feedback.

There isn’t a deterministic stage-to-model pipeline out of the box yet, but the idea matches what we’re already discussing. The best way to amplify the signal is to add your vote and details here: # Feature Request: User-configurable model routing rules for the main agent

For Router updates, including expansion to other plans, it’s easiest to follow the main thread here: Introducing Cursor Router

If you have questions about subagents with model: frontmatter or Plan mode, ping me and we can dig into it.