Hi just curious, i have been trying Opus 4.7 for sometime now and i just noticed something, it seems Opus 4.7 is still not integrated or optimized enough with the cursor SubAgent ? i mean when i’m using Opus 4.6 my context windows somehow “feel” more manageable because 4.6 using like 3-5 subagent without me asking in planning step to understand my codebase troughly, and its actually works great for me cause the planning result to execution often still less than 200k contex windows. but with opus 4.7 it seems the model trying to handle thing on his own if i dont specified to use the subagent, so in xhigh for planning alone i already spend 70% of my 300K context windows wich if i continue it will hallucinate in the middle of working it or made the result worse than Opus 4.6, so i decided to rollback to 4.6 , is anyone experiencing the same thing like me ? or is it just me ?
Hey, thanks for calling this out. The model deciding whether to call subagents during planning is model behavior, not a separate setting in Cursor. 4.7 and 4.6 have different defaults for how they do task decomposition, and what you’re seeing (4.6 decomposes into subagents more often, 4.7 tries to keep everything in the main context) has come up a few times in feedback.
A few ways to work around it:
- In your planning prompt, ask explicitly: “use subagents to explore the codebase before planning” or “spawn subagents to investigate X, Y, Z in parallel”. 4.7 follows this well if the request is clear.
- Create custom subagents for specific tasks (analysis, research, etc.). Then the model sees them as available tools and is more likely to use them. Docs: Subagents | Cursor Docs
- If you’re on xhigh, keep in mind reasoning effort uses context by itself, so xhigh plus a large request without decomposition can hit the 300k limit quickly.
For general feedback on 4.7, post it in the main thread: Opus 4.7 - Out Now!. The team is reading observations about model behavior there.