Lessons / Power
Subagents & parallelism
The biggest throughput ceiling for a power user is the single-threaded loop: ask, wait, ask, wait. Parallelism is how one person does the work of a team.
Instead of running tasks one after another, launch multiple agents at once — one to research, one to audit, one to draft — and synthesize when they all return. For writes, keep it to one agent per file (parallel writers collide); for reads, fan out as wide as you want.
A clean handle for which work fans out: prep parallelizes, acting sequences. The read-heavy prep — research, gathering, auditing — is independent, so run it wide. The write-heavy acting — anything that changes shared state — wants one lane at a time so two agents don't clobber each other. Fan out the prep, converge, then act.
Two honest caveats sharp operators catch fast. First: parallelism buys speed, not synthesis. Three agents fanned across three topics hand you three reports — coverage, not a decision. Real synthesis comes from choosing lanes that converge on one question ("what's the single throughline here?"), not lanes chosen for breadth. Second: if you're in a plain chat without true multi-agent infrastructure, the practical version of this is bundling several tasks into one message instead of feeding them one at a time — same intent (stop waiting in a single-file line), no API required.
Try it now
Take a task with independent parts and run them concurrently:
Launch separate agents in parallel: one to research X, one to audit Y, one to draft Z. Don't wait for each to finish before starting the next — run them concurrently and give me a synthesized result when they're all back.
You've got it when…
Independent work ran concurrently — not sequentially — and came back synthesized, not just concatenated. Wall-clock time dropped to the length of the slowest single task, not the sum of all of them.
Quiz — did it land?
Your tutor checks these before marking the lesson complete:
- Why does fanning work out to parallel agents beat the ask-wait-ask loop?
- Run 3 independent tasks concurrently — was the result genuinely synthesized, or just concatenated?