the problems
lost contextevery session starts from scratch
no parallelismresearch, plan, code: strictly one at a time
throwaway planyou burn through it, then it's gone
how far did we get?plan kicked off, is it done? no one knows
model picked by handyou choose which to run, every time
how i fixed them
shared contextevery agent knows where we left off
workflow in subagentsthe task splits, scouts run in parallel
plan always on handopen the plan whenever you want
sequential clustersthey run in order, you see what's closed
model nudgeheavy task or light? it suggests the right model