Build
The
Crew.
09.29.26
Your agent holds
four jobs.
Hire three more.
You handed one agent a whole job: research, write, check, send. The draft came back fast and clean, with two wrong numbers and nothing flagging either one. This issue shows you how to split the work across a small crew of narrow agents, choose how work moves between them, and stay in the coordinator seat until the handoffs prove themselves.
"After This Video, You'll Actually Understand Agent Orchestration." youtube.com/watch?v=PJnAPp_E6R8. A three-minute explainer covering what an agent is, why one agent struggles on big jobs, the orchestrator role, and three handoff patterns.
Executive summary →
Eight takeaways from the source, plus the three patterns drawn out.
New video concept →
Core message, opening premise, five teaching points for solo builders.
The crew, built →
A worked example: the Friday client update as a three-agent chain.
Counterpoint and your move →
What the video skips, and a six-step challenge for this week.
What the video
teaches
An agent acts. A chatbot answers.
The video draws one line first. A chatbot returns an answer. An agent takes actions toward a task: uses tools, searches the web, writes files, calls other software, checks its own work, and picks the next step based on what it finds. Every later idea in the video rests on this definition.
One agent breaks on big, messy jobs.
A single agent handles a lot. Hand one a job with many moving parts and quality slips. The agent plans, researches, writes, and checks at the same time, and details fall into the gaps between those roles. The video names this failure as the problem orchestration solves.
Orchestration means a team of narrow agents.
Agent orchestration means several agents working together, each owning a smaller, clearer piece of the job. The example: one agent researches, a second writes, a third checks the writing for mistakes. A coordinating layer assembles the pieces in the right order.
The orchestrator manages. The orchestrator does no work.
The coordinator goes by the name orchestrator. The video frames the orchestrator as a manager: the orchestrator decides what happens, who handles each piece, and in what order. Research, writing, and checking stay with the specialists.
Two reasons to split the job.
First, focus. A small agent with one job performs better than a big agent juggling every job at once. Second, diagnosis. When something goes wrong, you see which step broke, instead of untangling one long, knotted process.
Three ways to pass work.
In order: agent one hands results to agent two, agent two hands results to agent three. Use a chain when each step needs the step before. At the same time: independent tasks run in parallel and merge at the end. Manager routes: incoming requests get sorted and sent to the specialist built for that type.
Ask: one job, or several stitched together?
Before building a research tool, a support system, or a content pipeline, ask the video's one question. If the answer is several jobs, stop forcing one agent to carry all of them. Split the work into narrow agents with clear purposes and add a simple coordinator on top.
Skip building the plumbing yourself.
Zapier handles much of the connecting and handoff work between tools and agents. Most tools for building with Claude, ChatGPT, or Claude Code support this setup. The video closes on one line: a team of focused agents instead of one agent doing everyone's job.
Split big jobs into narrow roles. Pick how work moves between them. Put one coordinator on top. When a result goes wrong, the split tells you where.
Three ways
work moves
A. In order
Each step needs the one before. Slowest, easiest to debug.
B. Same time
Tasks ignore each other. Fastest, needs a merge step.
C. Routed
Work arrives in different shapes. Sort first, then send.
Stop hiring
one genius.
Build the
crew.
Your agent isn't weak. Your agent holds four jobs at once and grades its own work. Count the jobs, give each one an agent with one output, choose how work moves between them, and sit in the coordinator seat yourself until the handoffs earn automation. You get fewer silent misses, and when something breaks, you know where.
Fast. Clean.
Wrong.
You gave one agent the whole job. "Research my three competitors, write the client update, check the numbers, send me the draft." The draft came back in two minutes and read well. Two numbers were wrong. One competitor was missing. Nothing in the draft flagged either problem.
So you blamed the model. You tried a newer one and got the same shape of mistake. The model was never the problem. You hired one worker for four roles and asked the worker to grade their own work.
Agencies solve this with people: a researcher, a writer, an editor, a producer who keeps them in order. You work alone. You have no editor, no producer, and no budget for either. Orchestration gives you the structure of a team without the payroll.
Read your last prompt out loud. How many verbs did you hand one agent?
Count the jobs
before you build
Every request hides a job count. Circle the verbs. "Research, compare, write, check, send" holds five jobs. The source video's test asks one thing: one job, or several stitched together?
"Pull last week's analytics, compare against the week before, write a summary for the client, and make sure the numbers are right." Four verbs: pull, compare, write, check. Four jobs. One agent doing all four explains the wrong numbers.
Nouns hide jobs. "A polished report" hides edit. "An accurate summary" hides verify. "A client-ready draft" hides format and proofread. Rewrite the request with verbs only, then count.
Open your three most-used prompts. Circle every verb, including the hidden ones. Write the count in the margin. Any prompt over three is your first crew.
One agent.
One job. One output.
A narrow agent wins because its instructions stop competing. The writer stops worrying about accuracy. The checker stops worrying about tone. Give every agent a role card with five lines, and keep each line to one sentence.
JOB: [one verb, one object] Example: Check every number in the draft against the source sheet. INPUT: [what arrives, and from which agent] Example: The writer's draft and the researcher's facts table. OUTPUT: [one artifact, one format] Example: PASS, or a numbered list of mismatches with the correct value. DONE WHEN: [a test you run without asking the agent] Example: Every number in the draft appears in the facts table. NEVER: [the job belonging to another agent] Example: Never rewrite sentences. Never add new facts.
Without the NEVER line, roles drift. The checker starts rewriting, the writer starts researching, and you rebuild the one-agent mess with extra steps. The NEVER line keeps each agent in its lane.
Pick the handoff
by the dependency
The video's three patterns each answer one question. Ask the questions in order and stop at the first yes.
Chain them.
Research, then write, then check. Each step leaves a file you read, so a broken result points to one step.
Fan out, then merge.
Three competitor profiles, one agent each, one merged table at the end. Faster, but the merge step needs its own role card.
Route first.
Inbox triage: invoice questions to the billing agent, scheduling to the calendar agent, new leads to intake. The router only sorts. The router never answers.
Start with the chain. Chains run slower, but you read every step's output. Move to fan-out or routing once the chain runs clean for a week.
The handoff
is a document
Most crews fail at the seam, not inside an agent. Agent two sees only what agent one passes along. A vague handoff like "here's my research" strips out the sources, the doubts, and everything agent one threw away.
FROM: [agent] TO: [agent] WHAT I PRODUCED: [the artifact, attached or pasted below] SOURCES USED: [a link or file name for every claim] WHAT I'M UNSURE ABOUT: [each doubt, one line apiece] WHAT I LEFT OUT AND WHY: [anything cut, with the reason]
The checker never sees the writer's instructions. A checker sharing the writer's goals shares the writer's blind spots. Hand the checker two things: the source and the draft. Nothing else.
Save each handoff as a file with a date in the name. When the Friday update goes wrong, you open three files and find the step where the error first appears, instead of rerunning the whole job and hoping.
Here's my research on the three competitors. Pricing looks similar across the board.
Produced: pricing table, 3 rows. Sources: each vendor's pricing page, dated today. Unsure: Vendor B hides enterprise pricing. Left out: Vendor D. Page returned an error.
Be the orchestrator
first
The video puts a simple coordinator on top. For your first crew, you are the coordinator. Automation comes after you know which handoffs need a human.
Decide what happens when the checker rejects the draft. One reject sends the draft back to the writer with the mismatch list. Two rejects stop the crew and send the job to you. Without a stop rule, agents loop, and you pay for every lap.
A handoff earns automation when your log shows zero fixes for five straight runs. Not before.
MON researcher → writer fixed: missing source on revenue row TUE writer → checker no fix WED researcher → writer fixed: pulled last month, not last week THU writer → checker no fix FRI writer → checker no fix → automate this handoff first
The Friday client
update, as a crew
Before: one prompt asking for research, a draft, and a number check. Then a round of fixes by you. After: three narrow agents in a chain, with you in the coordinator seat.
Researcher
Pulls this week's metrics from the client sheet and decisions from the call notes. Output: a facts table with a source on every row. Never writes prose.
Writer
Turns the facts table into a 200-word update in your voice. Never adds a number missing from the table.
Checker
Sees the facts table and the draft, nothing else. Output: PASS, or a numbered list of mismatches. Never rewrites.
More agents,
more places to break
Errors multiply across steps.
Three agents at 90% each finish clean about 73% of the time (0.9 × 0.9 × 0.9). Every added agent needs a checker catching its misses, or the crew performs worse than one agent.
Every agent costs a call.
Three agents means three model calls at minimum, plus retries. Splitting a two-verb job raises the bill and the wait for no gain in quality.
Context dies at the seam.
The writer never sees what the researcher threw away. Without a handoff format, the crew knows less than the single agent did.
No plan for failure.
The video never says what happens when the checker rejects the draft. Without a retry limit, agents loop and burn tokens.
Focus is a rule of thumb.
Current single agents handle long tasks well. Split when you see a specific miss, not because a crew diagram looks professional.
Your move.
Pick one job an agent got wrong this month. Build the crew for that job this week.
Build the crew.
Keep the call.
09.29.26