Agent launches expensive Task subagents / high-tier models without explicit user consent

Support Bug Reports
,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Summary:
In Agent mode, the model repeatedly invoked the Task tool with high-cost model slugs (e.g. claude-sonnet-5-thinking-high or GPT-5.5-high) to implement a large multi-milestone plan. The user did not request subagents or premium models. This caused high API spend, an API usage limit switch, and a long-running subagent that had to be interrupted. The user then had to revert the work.

Expected:

  • Default to the parent/session model unless the user asks for a specific model or subagent.
  • Do not auto-launch expensive parallel/background agents for large implementation plans.
  • Prefer in-session incremental edits for coding playbooks.
  • Surface cost/model choice before spawning Task agents on premium models.

Actual:

  • Agent chose Task + premium thinking models on its own.
  • Failed model slug → retry → API limit → model switch → another long Task run.
  • User could not safely “leave it running.”

Impact: Unexpected high spend, interrupted work, full revert of implementation.

Steps to Reproduce

  1. Large attached plan with many todos (“implement the plan… don’t stop until all todos”).
  2. Agent mode.
  3. Observe Task tool calls with expensive model values without user requesting them.
  4. Explorer model is set to Composer 2.5, but that was not exploration, it was implementing stuff through expensive sub agents.

Operating System

Linux

Version Information

Version: 3.10.20
VS Code Extension API: 1.125.0
Commit: 23b9fb205fe595ea2be29da7214e19762d037fc0
Date: 2026-07-07T07:03:33.071Z
Layout: IDE
Build Type: Stable
Release Track: Default
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.256
OS: Linux x64 6.8.0-134-generic

For AI issues: which model did you use?

Cursor Grok 4.5

Additional Information

It’s scary to implement anything with Cursor like this, the agents cannot be trusted.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

6

Hey there,
On your plan, subagents are meant to run on your session model, and the agent reaching for premium models like Opus and Sonnet on its own for Task subagents is a known issue our team is working to improve.

One clarification that trips people up: the Explorer model setting (Composer 2.5) only controls the built-in Explore (codebase-search) subagent. It doesn’t govern the general subagents the agent spins up to implement a plan, which is why those ran on other models despite that setting.

Until the improvement lands, these are the most reliable ways to keep it in check:

  1. Cap your spend. Set an on-demand spend limit in your dashboard’s Spending tab. When it’s reached, AI features pause instead of running up more usage.
  2. Turn off the models you don’t want. In Cursor Settings → Models, disable the premium models (Opus, Sonnet, etc.) so the agent can’t pick them for subagents.
  3. Use Auto as your main model. Subagents follow the parent model, and Auto is cost-managed, so this keeps subagent spend down.
  4. Plan big work first. For large multi-milestone tasks, use Plan mode and review before you hit Build, rather than one “implement everything, don’t stop” prompt. You can also add a project rule in .cursor/rules telling the agent not to spawn subagents or switch models (Rules).

To recover the interrupted work, you can roll back with Restore Checkpoint / Revert on an earlier message (details).

If this ate into your usage unexpectedly, email [email protected] with the request ID and we can take a look at your account.

8

Thank you Mohit!