Noticed that when you ask Composer 2.5 to work in parallel inside a plan, it doesn’t actually spawn multiple agents — it just runs everything through one agent sequentially. Kind of defeats the purpose of “parallel” in the plan.
Tried adding this to the plan prompt to force it to actually split the work:
Group the tasks into batches that can be executed in parallel. If any tasks
must be done sequentially, indicate that explicitly (dependencies).
In the plan, note that for multi-task execution we will spawn sequential agents
(where possible) to finish tasks that parallel agents may depend on later.
All spawned agents must inherit the same model as the planning agent.
Haven’t set this up as a .cursor/rules file or a proper subagent file (.cursor/agents/*.md) yet — just tested it as a raw prompt in the plan. Curious if anyone’s already automated this via rules/subagents and whether it actually improves parallelism, or if Composer still collapses it into one agent regardless.