Subagents: stop running Claude like one assistant

Most people run Claude like one overworked assistant and wonder why hour three is worse than hour one.

Every file it reads stays in the window,
Every grep, every log dump, every wrong turn,
Until it's answering you through a fog of half-relevant code.

Subagents fix this. Separate Claude instance, own context window, does the digging, sends back one paragraph.

The forty tool calls it burned? Never touch your window 🔥

And because they're separate, you can run four at once. Four researchers, four minutes, four summaries.

I wrote the whole playbook. Five copy-paste briefs, a working custom agent file, and the decision table for dispatch vs do it yourself.

Free. No email wall.

Comment "AGENTS" 👇 and I'll send it to your DMs.