compaction is much better now so for certain things you may not need to explicitly clear or externalize context to for the sake of model intelligence/focus. (still useful for collab w/ humans or handoff between agents / machines though!)
"dumb zone" a little more flexible for 1m frontier models. I usually my autocompact to 400k and tend to do a manual compact between 200-400k depending on model/harness. less so clearing now
3-tier subagent architectures make this a little better too. You can have a top level orchestrator that runs the process, and can launch subagents that can themselves launch subagents. So instead of human steering two levels, top level can.
i don't think there's any good reason to go beyond 3-levels of subagent nesting today for majority of tasks.
YMMV, as always these are guidelines / training wheels. your intuition as to whether model is on track or getting dumb is always the best source!