Feature Request: Dynamic model assignment for sub-agents (similar to Copilot)

Ideas Feature Requests
,
1

I’m a Cursor Enterprise user, and I’d like to advocate for the ability to dynamically specify sub-agent models. While I can achieve this in Copilot, I’m currently limited in Cursor.

Here’s a simple example: I use one sub-agent for implementation and then spawn three separate sub-agents for review—each using a different model: GPT-5.5, Gemini-3.1-Pro, and Opus-4.6. In my testing, I’ve found that it’s rare for all three models to pass the implementation on the first try, which highlights why model diversity is so critical for specific tasks.

I’m aware that I could manually create three separate sub-agent files and hardcode the model in each. However, that doesn’t scale. In my actual workflow, I might have nearly a hundred sub-agents. More importantly, the choice of model shouldn’t be hardcoded within the sub-agent itself; it should be dynamically decided by the Orchestrator model based on the context.

Cursor needs to support dynamic model assignment at the time of calling a sub-agent, similar to how Copilot handles it. This level of control is essential for complex, multi-agent orchestration.

4

One workaround is to use cursor-agent CLI to spawn subagents with specific model.

5

Fully agreed, choosing the model at call time is a very basic requirement. It sometimes depends on the complexity of the task that the subagent will do, and the current workaround is very cumbersome.

6

I’m not sure whether to be disappointed or relieved that Cursor lacks the option to specify a model when spawning a sub-agent. Once I realized I could use the CLI to spawn sub-agents with my own model choices, I began to wonder: why stay tethered to Cursor’s CLI at all? This prompted me to explore using Codex, Copilot, Claude, and Gemini directly.

The result was a total eye-opener. A multi-agent workflow that would have cost me $500 per run within the closed Cursor ecosystem now costs less than half of that. While I initially wanted this feature built into Cursor, their current limitation forced me to optimize my own stack—and the cost savings have been massive.