Sub agents make the most sense at pre existing boundaries that already have a high impedance mismatch for humans.
Without sub agents I would be able to test maybe two product scenarios before the main context needs compacting. With sub agents I can go 40-50 scenarios wide and easily compare their results in a single turn of the main agent loop. The loss in fidelity is a feature, not a bug in this context. I don't want the main agent to "see" literally every logical view that was experienced. I only want it to receive a testing summary with respect to the original instructions and experience log. Anything beyond this is too much, even for a human.
Without sub agents I would be able to test maybe two product scenarios before the main context needs compacting. With sub agents I can go 40-50 scenarios wide and easily compare their results in a single turn of the main agent loop. The loss in fidelity is a feature, not a bug in this context. I don't want the main agent to "see" literally every logical view that was experienced. I only want it to receive a testing summary with respect to the original instructions and experience log. Anything beyond this is too much, even for a human.
But a fork is prompt-cache friendly so it might end up costing lesser