Parallel Plates

Watch a busy kitchen at seven p.m.: three pans on fire, two plates being dressed, one cook tasting stock — simultaneously. Nobody waits for the fish to finish before starting the salad. That's not heroics; it's the entire reason kitchens have more than one cook: independent work happens at once.

Your crew can do this. Subagents don't share hands — three of them genuinely work in parallel while you (and your head chef) watch the pass. But the magic word is doing quiet, load-bearing work in that sentence: independent. Parallelize work whose parts touch each other and you get the multi-agent version of two cooks reaching into one pan. So today teaches the two halves as one skill: running parallel plates, and — the harder, more valuable half — carving jobs so the parts can't collide.

What you'll plate today

A real three-part job, carved and cooked by three subagents at once — plus the carving test you'll apply forever to know what parallelizes and what never should.

Ingredients

• Your kitchen-brigade project
• A real job with survey-shaped parts — three documents to summarize, three options to research, three sections of something to draft
• About 30 minutes

Cook

1. Learn the carving test first

Here's a job: [your real job — e.g. "assess these three vendor
quotes in quotes/ and recommend one"]. Before any cooking: carve
it. Split into parts and apply the INDEPENDENCE TEST to each pair:

1. Could cook B do their part WITHOUT seeing cook A's result?
2. If both touch the same file or fact, do they only READ it?
3. Does the ORDER of finishing matter?

Yes / yes / no = parallel-safe. Show me the carving and the test
results — and if any part fails, show which parts must wait for
which. Don't cook yet.

Study the carving — this is the lesson's real skill. "Summarize quote A / quote B / quote C" passes clean: nobody needs anybody. But "…and recommend one" fails question 1 — the recommender needs all three summaries. So the job carves into a parallel phase and a merge phase, and seeing that structure before cooking is what separates a brigade from a pile-up.

2. Fire three plates

Cook the parallel phase: three subagents at once, one quote each.
Each gets the same order — summarize per our handbook's format:
costs, terms, risks, one-line character read. They work from the
file only; no peeking at each other.

While they cook, narrate the pass for me: who's working, who's
done. Then show me all three plates side by side, unmerged.

Watch your terminal earn the course's name — multiple workers, one pass, you expediting. Then read the three summaries before any merge. Side-by-side raw output is your quality window: is one summary thinner? Did they interpret "risks" consistently? This glance is what the merge phase inherits, so taste now.

3. The merge — where parallel earns its keep

Now the merge phase, one cook: take the three summaries and build
the comparison — where they agree, where they differ, and the
recommendation with reasons. Then send the whole thing to The
Inspector before it reaches me.

Notice the shape of what just happened: parallel for the independent parts, sequential for the parts that need everything, inspection before the pass. That's not a trick for vendor quotes — it's the shape of almost every crew-worthy job you'll ever meet. Research-then-synthesize, draft-sections-then-unify, check-everything-then-summarize: same skeleton every time.

4. Put the pattern on the chart

Add to STATIONS.md under PATTERNS: "Parallel plates — for jobs
with independent parts. Carve with the independence test, fire
parallel, merge sequential, inspect before serving." And the
honest counterweight in CLAUDE.md: parallel costs more than solo
— three cooks burn three cooks' worth of my subscription. Use it
when the parts are truly independent AND the job is big enough
to feel the wait. Otherwise: one cook.

Save point: "parallel plates".

That counterweight isn't fine print — it's Lesson 9's opening argument arriving early. Parallel is faster, never cheaper. A three-minute job done by three cooks in one minute spent three minutes of capacity plus coordination. Feel-the-wait is the honest threshold.

When it burns

• Two cooks came back with contradictory takes on the same fact — not a bug, a finding: they read the same source differently, which is exactly what the merge phase is for. "Both readings, side by side, with the source line — I'll adjudicate." Parallel eyes catching ambiguity is a feature you paid for.
• One plate came back thin — same order, different diligence; it happens. Re-fire just that plate: "redo quote B's summary with the same order — match the depth of the other two." Re-firing one plate is cheap; that's another thing parallel buys you.
• You can't find a parallel-shaped job — then don't manufacture one; the counterweight rule is already doing its job. The shape appears naturally around surveys: multiple documents, options, candidates, weeks of journal. When it appears, you're ready.
• The merge cook ignored one summary — merge orders need the completeness clause: "the comparison must cite all three by name; note explicitly if one contributes nothing." Cold cooks do what orders say — Station Chart lesson, still true.
• It all worked and took longer than doing it solo — honest possibility for small jobs! Coordination has overhead; that's the feel-the-wait threshold teaching itself. Log the experience — Lesson 9 will price it properly.

Order up

□ A real job carved with the independence test — merge phase identified
□ Three plates fired at once; raw output tasted side by side
□ Merge phase cooked sequential, Inspector before the pass
□ The pattern and its honest counterweight are on the chart
□ Save point "parallel plates"

Next up — Lesson 06: The Expediter. Carving, firing, merging, inspecting — tonight you conducted it by hand. Next lesson that choreography becomes a written routine the kitchen runs from one order: the expediter's call sheet.

Stuck on a step? Question box below.