One Person, 8 AI Agents, 3,000 Baseball Caps in Two Months

Written with AI assistance. Figures without a traceable source were cut before publishing.

The same piece is posted as a thread on GitHub — that copy has a reply box under it, and this one does not.

Zhang Qianchao runs a custom baseball cap export business on Alibaba.com with 8 AI agents and no employees, and in 2 months he shipped 3,000 caps to buyers across Europe, the Americas and South America at roughly the throughput of a 9-person team. The numbers underneath that headline are the interesting part: 3 minutes for a deep market research pass, 1,800 designs produced in a month, 12 minutes from a customer request to a rendering, and orders signed on the spot while the buyer was still in the conversation. I think the last one is the actual mechanism, and almost nobody talks about it.

Speed Is a Closing Tactic

Twelve minutes to a rendering does not sound like a business model. It is one.

In custom manufacturing the normal sequence is that a buyer describes what they want, the supplier goes away, a designer produces something over a day or three, and the buyer looks at it after their enthusiasm and their attention have both moved on. Every hour in that gap is an hour the buyer spends talking to someone else. Compressing product selection and design into a single day removes most of that gap, and compressing the rendering itself to 12 minutes removes it entirely, because the buyer is still on the call when the picture arrives.

Signed on the spot. That phrase is doing the work here.

I’d argue this is the most transferable lesson in the whole case, and it has nothing to do with the specific tools. Anywhere a sale requires a custom artifact before the customer can commit, the time to produce that artifact is not an operational cost, it is a conversion variable. Most businesses classify it under production and optimize it for cost.

Eight Agents Means a Relay, Not a Fleet

The design detail that makes 8 agents workable rather than chaotic is that they hand off to each other automatically.

That sounds mundane and it is the difference between a system and a pile of chatbots. Each agent covers a stage — research, selection, design, customer background checks, and so on — and the output of one becomes the input of the next without a person copying anything between windows. Manual handoff is where most single-operator setups quietly fail, because the human becomes the message bus and the whole arrangement runs at human speed while costing agent money.

What the relay buys you is compression at the seams, not raw capability at any single step. 1,800 designs in a month is impressive; a design pipeline that starts the moment research finishes is what makes 1,800 designs reach anyone.

The asset structure is worth stating plainly, since it is the part that makes this reproducible: zero inventory, zero design team, zero sales staff. Extremely light. Nothing in that model requires capital, which is why it is available to an individual at all, and also why it will not stay uncrowded.

The Constraint Nobody Mentions

Here is the line from Zhang’s own account that I keep returning to: AI amplifies the cognition you already have, it does not hand you cognition out of nowhere.

Read against the rest of the case, that is a fairly severe entry requirement. This model suits someone with an existing foreign-trade or e-commerce background who is willing to learn AI tooling. The 3-minute research pass produces a document; knowing which parts of it are wrong is a separate skill that came from years of doing the job manually. The 1,800 designs need somebody who can look at a batch and know which ones a buyer in a specific market will actually pay for.

Domain knowledge first. Tools second.

I suspect that is why so many attempts to copy cases like this one fail without producing an obvious failure. The agents run, output appears, nothing errors, and the operator has no way to tell that they are shipping the wrong products very efficiently. Honestly, if I were advising someone with no trade background who wanted to run this playbook, I would tell them to go get the background, which is not the answer anyone wants.

What I Would Verify Before Copying It

Every figure here comes from one operator describing his own results, and none of it is audited. That does not make it false; it means the useful move is to identify which numbers you could reproduce cheaply and test those first.

Start with a business metric rather than an efficiency claim. AI project value cannot stop at “we saved time” — it has to name whose efficiency improved, where the saved hours went, whether rework fell, whether conversion rose, whether service cost dropped. In this case the metric that matters is close-rate on live conversations, and that one is measurable in a week without building anything at all. If sending a rendering during the call does not move it in your category, the other 7 agents will not save you.

Give the direction three days, not three months. Spending more than about two weeks choosing what to sell is research being used as procrastination, and the market answers in half a month what a solo researcher cannot settle in half a year. Try to sell 3 of something before the automation exists.

And keep the trust layer human and real. Overseas buyers evaluating a supplier want to see the factory, the production line, actual finished cases — the concrete evidence that you exist and can deliver. Polished promotional video is not what closes that gap. Generated imagery is not either, and I would be careful about how much of the customer-facing surface ends up machine-produced, because the thing being evaluated is precisely whether there is anything real behind the listing.

Where the Fragility Sits

Two failure modes deserve a mention, because neither shows up in a case study written by the person who succeeded.

The first is platform dependence. This entire operation lives on Alibaba.com and its Accio Work platform, which means the storefront, the discovery and the buyer relationship are all rented. That is a reasonable trade for someone with no capital, and it is still a rented position; a policy change reprices your business without asking. The version of this I have seen bite hardest is on the payments side, where one developer’s Stripe account was frozen for a dispute rate above 1% and 1,500-plus paying subscriptions stopped at once. The product was fine. The business stopped anyway. Spread the collection channels, build some way to reach your buyers that the platform does not own, and watch the metrics that predict enforcement rather than the ones that flatter you.

The second is that a relay hides its own errors. When 8 agents pass work forward automatically, a bad research pass propagates into selection, into design, into the buyer conversation, and by the time anyone notices, the mistake is downstream of 4 steps that each looked like they worked. The fix is boring: put an acceptance check at the seam where the cost of being wrong jumps, which here is the moment a design goes in front of a customer. One check, at the expensive boundary. Not eight.

The Short Version

A one-person operation matching a 9-person team is a real result and a badly framed one, because the comparison implies the AI replaced 8 people. It did not. It removed the waiting between steps that 9 people would have spent coordinating with each other, and it did that for someone who already knew which decisions to make at each step.

If you have the domain knowledge, the relay structure is worth building and the seams are where the gain lives. If you do not, 8 agents will get you to the wrong answer in 3 minutes instead of an afternoon, and the invoice arrives either way.

Also readable on Telegraph.

Read next

All 49 write-ups

The 7 figures in this piece — each with the sentence it came from — are in the figures table, alongside 401 more, as JSON and CSV.

Topics: Automation Systems · Niche Market · Revenue Growth

Part of llm-api-pricing — field notes on AI coding agents.

Did this save you an afternoon? A star on the repository is the whole ask — it is what puts these in front of the next person looking; the data is CC BY and does not require starring.

One thing this piece could not settle: How long does it take you today to get a custom mockup back in front of a buyer who is still in the conversation? Reply with a number of minutes. The reply box is on the thread copy of this piece.

Want a figure that is not in here yet? Say which metric, which provider, which unit — in one line. One required field, and the page you came from is already filled in. Got a better number? Open an issue — that form knows which write-up you came from too; corrections and counter-data are the point.