I just signed up again to try out Composer 2.5, but noticed something when using it. I usually give my agents instructions and very often have “oh but also do this …” moments where i drop in more messages to streer the work of the agents. Works fine in Claude code and codex, but in Cursor, it just does not work that way, it just drops anything it’s currently doing and starts doing the stuff in the new message. This is absolutely infuriating!
Steps to Reproduce
Tell it to do A first, then also add “do B too”, and it will drop A and do B, leaving work on A open…
Expected Behavior
“Alright, after I do A, I will do B” or “while I do A, I will let sub agent do B” → like any other cli
Hey, thanks for the report. This is a known limitation in the CLI: follow-up messages are currently treated as a new turn, not as a steering hint on top of the current task. The CLI does have a queue, but when it processes a queued message, the agent still switches to it and drops what it was doing. There isn’t an equivalent of Claude Code’s /btw yet, and I can’t give an ETA.
If most of your work is in the IDE and not strictly in the CLI, take a look at multitask, like @Tom_Coustols mentioned above. The idea there is that a parent agent delegates parallel tasks to sub-agents, and follow-up messages behave more logically. It’s not a 1-to-1 replacement for /btw, but for the use case “do A, and also B” it’s a closer fit.
For the CLI, the current workaround is to wait until the current turn finishes before adding a new instruction, or to include everything in the first prompt like “do A, then B.” I know that’s not what you want. I’ll pass the feedback to the team, this pattern is pretty standard in competing tools.
/btw is not what I am looking for, that is a separate thread. I need steering, and I also don’t need the IDE since I am not necessarily working on a coding project. I run CLI’s on various systems, and I am very often just in a root folder and not a project folder with code.