When a user configures Cursor Settings → Agents → Subagents → “Explore subagent model” to Composer 2.5, the parent (foreground) agent still launches background Task subagents with an explicit model: claude-4.6-opus-high-thinking parameter, ignoring the user’s subagent model preference.
Impact
With **Composer 2.5** configured under **Settings → Agents → Subagents**, the parent agent launched two parallel background subagents for delivery-loop work on issues **#116** and **#117**. Both ran as **Opus 4.6 High** instead of the configured subagent model. Running two expensive premium-model workers in parallel drove API usage to **100%** — a significant, unexpected cost/usage spike the user did not intend when they set subagents to Composer 2.5.
Steps to Reproduce
Open Cursor Settings → Agents → Subagents
Set “Explore subagent model” to Composer 2.5 (see screenshot below)
In Multitask Mode, ask the parent agent to delegate work to background subagents (e.g. "run agent-delivery-loop for issue#116and#117in two subagents")
Observe the parent agent launches Task subagents with model: claude-4.6-opus-high-thinking explicitly set in the tool call
Expected behavior
Background subagents should use the user’s configured subagent model (Composer 2.5) from Settings
Parent agent should not override model selection unless the user explicitly requests a specific model in chat
When no model is specified by the user, the Task tool should inherit/respect Settings → Agents → Subagents rather than the parent agent picking the strongest available model
Actual behavior
Parent agent autonomously selects claude-4.6-opus-high-thinking for generalPurpose Task subagents
User’s Composer 2.5 subagent setting is ignored
User pays higher cost / different latency than configured
For AI issues: add Request ID with privacy disabled
4f623943-4bed-4e0c-b51a-60f32cf66114
Additional Information
This is recurring, not a one-off: with Composer 2.5 configured under Settings → Agents → Subagents, parallel background subagents spawned via the Task tool have repeatedly included an explicit model field even though the user did not request a specific model in chat. Observed overrides include Opus 4.6 (claude-4.6-opus-high-thinking) and Opus 4.8 (claude-opus-4-8-thinking-high or similar).
Hey, thanks for the detailed report with screenshots and the Request ID. Super helpful.
This is a known bug. The Explore subagent model setting currently doesn’t actually apply to Task subagents. In some cases the parent agent injects an explicit model when it starts background workers, in your case claude-4.6-opus-high-thinking, so your Composer 2.5 selection gets ignored. I checked the Request ID and the backend resolved the model correctly, so the issue is specifically that the parent agent is passing the model in the Task call itself.
We’re already tracking this on our side and I’ll link your thread to the existing issue. I can’t share an ETA for a fix yet. For context, here’s another thread with the same symptom: Explore subagent model doesn't work
Until this is fixed, to avoid the same kind of spike, in Multitask Mode it’s best to avoid running multiple background workers in parallel for heavy tasks, or explicitly specify the model in the prompt.
Also, if you want to discuss unintended API budget usage because of this, email [email protected] and they can review your account. I’ll post an update here as soon as we have one.
Hey @RaitP1, thanks for checking in. This is the same known bug described above. The Explore subagent model setting currently doesn’t apply to Task subagents, and in some cases the parent agent passes an explicit model, so your choice gets ignored. We’re tracking the issue, but I can’t share an exact fix date yet.
To clarify, are you using the IDE or the CLI, and what Cursor version are you on? There was another thread with similar but not identical behavior in the CLI, and I want to confirm this is the same case.
Until this is fixed, to avoid usage spikes, in Multitask Mode it’s better not to run multiple background workers in parallel on heavy tasks, or explicitly specify the model you want directly in the prompt. I’ll post an update in the thread when I have one.
Subagent frontmatter: name: researcher description: “Multi-source deep research agent. Searches the web, synthesizes findings, and delivers cited reports with source attribution. Use when thorough research on any topic with evidence and citations is needed.” model: composer-2.5-fast
Parent chat tool calls:
{“type”:“tool_use”,“name”:“Task”,“input”:{
“description”:“Research skill composition patterns”,
“prompt”:“Research industry best practices for AI agent skill composition…”,
“subagent_type”:“researcher”,
“model”:“claude-4.6-sonnet-medium-thinking”
}}
{“type”:“tool_use”,“name”:“Task”,“input”:{
“description”:“Research config single source of truth”,
“prompt”:“Research best practices for maintaining single source of truth…”,
“subagent_type”:“researcher”,
“model”:“claude-4.6-sonnet-medium-thinking”
}}
{“type”:“tool_use”,“name”:“Task”,“input”:{
“description”:“Research agent skill deduplication”,
“prompt”:“Research how agent authoring toolkits typically relate to hand-maintained authoring guides/skills…”,
“subagent_type”:“researcher”,
“model”:“claude-4.6-sonnet-medium-thinking”
}}
@RaitP1, thanks for the details, they helped confirm it. This is the same known bug described above, and in your case it’s on the IDE side 3.7.19, not the CLI.
From your log, it’s the exact same pattern. The researcher sub-agent is configured with model: composer-2.5-fast, but the parent agent injects an explicit model: claude-4.6-sonnet-medium-thinking into the Task calls, so your setting gets ignored. In other words, the model is being overridden at the parent agent’s Task call level.
We’re tracking the issue, but I can’t share an exact fix date yet.
Until it’s fixed, to avoid usage spikes in Multitask Mode, it’s best not to run multiple background workers in parallel on heavy tasks, or explicitly specify the model you want directly in the prompt. I’ll post an update in the thread when there’s news.
