Cursor rules and sub-agent calls do not work when agent model is set to Auto or Composer 1

Describe the Bug

When the agent model is set to Auto, Cursor rules and sub-agent calls do not work as expected.

Specifically:

  • Sub-agents fail to execute because the agent responds that it cannot call a Task handle.
  • Cursor rules are not applied at all, even when alwaysApply: true is set.
  • Interestingly, AGENTS.md located at the project root is always loaded correctly, even though it does not specify alwaysApply: true. It is just a plain text file that lists a few simple rules.

This behavior is the same even when the model is manually set to Composer 1.

However, the issue does not occur with other models:

  • More expensive models such as GPT 5.2-Codex or Opus 4.6 work correctly.

  • Even cheaper models like Gemini 2.5 Flash work without any issues.

    • Sub-agent calls function properly.
    • Rules are applied as expected.

I am not sure whether this behavior is intended or a bug. However, considering that until about three days ago, rules were applied correctly even if sub-agent calls were unreliable, this seems likely to be a regression or bug specific to the Composer 1 model.

I would appreciate clarification on whether this is expected behavior or an unintended bug.

Steps to Reproduce

Set agent to “auto” or “Composer 1”

Expected Behavior

Rules are applied to agents and sub agent can be executed as expected

Operating System

Windows 10/11

Version Information

Cursor 2.4.28

For AI issues: add Request ID with privacy disabled

914daca9-bd97-4de1-9a8e-771f78416d32

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey, thanks for the report. This is a known issue. Composer 1 (and Auto, which uses it) doesn’t support sub-agents yet. The Task tool isn’t available for this model.

The team is aware, but there’s no timeline yet for when this will work.

Workaround (as you already noticed): use other models. GPT 5.2-Codex, Opus 4.6, or even Gemini 3 Flash work fine with sub-agents.

About rules, they should work with Composer 1. If rules don’t apply even without sub-agents, that might be a separate bug. Can you clarify: do rules not work ONLY when you try to call a sub-agent, or do they not work in all Composer 1 scenarios?

Thank you for the response, Dean.

The issue with Rules appears to be identical to the sub-agent invocation bug, including the conditions under which it occurs.

I have a rule file located at ~/.cursor/rules/absolute-rules.mdc with the following header configuration:

---
alwaysApply: true
description: MUST-FOLLOW RULES
globs: ["*"]
---

When I ask an agent using the Composer 1 model, “What rules are you currently following?”, the agent responds that it is only following the system-provided default rules and that there are no other rules applied.

Even after assigning the agent multiple tasks and asking the same question again, the agent never mentions absolute-rules.mdc. Additionally, when hovering over the context meter at the bottom of the agent chat panel, absolute-rules.mdc does not appear in the “Active Rules” list.

Notably, this problem does not occur when the same agent is loaded with a different model. The issue is reproducible only with Composer 1.

One important point is that until February 3, rule application was working correctly. I currently have around 15 rule files in my rules folder, and all of them were being applied exactly as intended before.

At the moment, the only rule file that consistently appears in the Active Rules list and is reliably applied is a single AGENTS.md file located at the project root. All other rules in the rules directory are mostly being ignored. Composer 1 agents are behaving in ways that don’t match my rules at all.

Thanks for clarifying, that’s a helpful detail.

So, for the two issues:

Sub-agents: Composer 1 doesn’t support the Task tool yet. That’s a current model limitation. The workaround is what you already found, use other models.

Rules: This one’s interesting and looks like a regression. If rules with alwaysApply: true worked in Composer 1 before Feb 3, but now they don’t show up in Active Rules, that’s a bug, not expected behavior. I’ll pass it to the team.

For now, the recommendation is the same: use other models for tasks where you need rules and sub-agents. I’ll update the thread if there’s news about rules.

I also notice that GPT-5.x is very inconsistent in starting subagents.

Often it will start doing the work subagents where supposed to do instead of starting subagents.