← Back to Blog

guide

By ReclaimLLM Team

Which model saw your code? How to audit Cursor subagent models

Learn why a Cursor subagent can use a different model from the parent chat and how to check task cards, usage records, and model controls.

A Cursor user reported selecting one model to build a plan and seeing most of the resulting consumption attributed to another. In a separate session reviewed in the same thread, staff found that the main chat stayed on the selected model while child agents handled delegated tasks on a different one. A session can therefore keep the original model in the parent chat and still generate usage on another model. User report · Staff session review

To check what ran, start with the subagent task card and the matching Usage dashboard row. The parent picker records the choice for the parent chat. For security review, you also need to establish which work a subagent received and whether it could access files in the workspace. Cursor subagent documentation

How Cursor chooses a subagent model

Cursor documents several paths. A built-in helper can select a model for its job. A custom subagent can inherit the parent's model or request a specific model through its model field. The parent agent can also name a model when it starts a subagent. Plan eligibility and team model restrictions can override a configuration, so verify the effective behavior in the team's environment. Cursor subagent documentation

The distinction matters on Auto, too. Cursor documents that a third-party subagent can generate third-party usage even when the parent uses Auto. On eligible legacy plans that can appear in the Other Models pool. The pool helps explain the charge, while the individual row helps identify the model. Cursor usage pools

What the support cases add

These reports describe particular environments and versions. Use them to decide what to test.

CaseWhat staff reported
Delegated plan tasksA parent stayed on its selected model while child agents ran on a different one. Session review
Plugin-packaged agentsIn the reported setup, plugin agents inherited the parent's model even with a model declared in the plugin agent file. Staff pointed to repo-local or user-local agent files as the workaround. Plugin report
Cloud UI testingStaff attributed Claude Sonnet usage to an autonomous UI-testing helper in the reported cloud workflow. Cloud agent report
Fast variant and its displayStaff found that automatic follow-up turns after background commands could use a Fast variant. They also identified a separate picker display issue where the default variant could appear even when the turn ran on another variant. Follow-up turns · Picker display
A chat using an Anthropic API keyStaff said a Claude parent running on the user's own Anthropic key kept its subagents on the parent model in that setup, overriding the custom model setting. API key report

A Fast variant can change usage without changing the provider. A subagent on a different model can change the provider, depending on the model and its serving arrangement. For an approval review, record both the model and any independently verified provider or endpoint information. A model name alone cannot show where processing took place or which source files were sent.

Verify a workflow before it sees sensitive code

  1. Pick a representative task. Use a disposable repository with synthetic content. Run the same local or cloud workflow the team intends to approve, including any built-in helper, plugin agent, or custom subagent.
  2. Capture the requested configuration. Record the parent model, each custom agent's model setting, the plan and client version, and whether the run uses an organization-managed or a personal key. Record key ownership without collecting key values.
  3. Compare the task card with the usage row. Cursor documents that the task card and Usage dashboard report the executed subagent model. Match rows to the run with available identifiers or timestamps. Where the evidence is ambiguous, preserve that uncertainty and narrow the test to one subagent at a time. Cursor subagent documentation
  4. Test the team restriction. Use the admin model settings to restrict the available models, then repeat the run. Check how the workflow behaves when it requests a disallowed model, including whether it stops or uses an approved fallback. Repeat after changes to the plan, plugin, or model controls. Cursor model controls

Keep the result with the team's rollout record. Include the subagent's task and its permissions so a later reviewer can see what it could access. If you need proof of what data a provider processed, use request evidence from the relevant telemetry or provider records where available. A usage row is evidence of model usage, with a narrower scope.

The Cursor security checklist covers the workspace settings, hooks, and MCP review that complement this check when onboarding a team.

FAQ

Do Cursor subagents use the parent model?

They can inherit it. Built-in helpers, custom configuration, and explicit model requests can produce a different choice, subject to plan and team controls. Cursor subagent documentation

Does bringing an API key change subagent behavior?

It can. Cursor staff documented inheritance behavior for a Claude chat using the user's own Anthropic key. Confirm with a small run before relying on it for another provider or key setup. Cursor staff reply

Which model should I record if the picker disagrees with usage?

Record the requested model from the picker and use the task card and matching usage row for execution evidence. Staff have documented a case where the picker showed the default variant. Cursor staff reply