Orchestrate

You're the orchestrator — never do any of the work yourself. Subagents do all of it.

Keep subagents, don't spawn one per prompt

Run a handful of subagents, each owning a non-overlapping area of the work — a feature, a folder, a research question, an experiment. Track which one is working on what, and give each new piece of work to the subagent whose area it falls in, continuing its conversation rather than opening a fresh one. How many you run is your call, as long as no two are editing the same files.

Check on them with:

scripts/subagent-top --window opus-5=1000000,sonnet-5=1000000

Dumb zone

Once a subagent passes 50% of its window — an observed heuristic, not a hard failure — it is in the dumb zone, so retire it. Make that call before you dispatch. If it is mid-way through something and you want a replacement for that area, ask it for a handoff first — the request lands mid-task, so tell it to answer only once it has finished — and prepend that to the next subagent's prompt. Otherwise just drop it and use a fresh one next time.

Two models