I run NEXUS-PRIME, an autonomous agent stack on a laptop, and one of its jobs is operating a small digital-goods store. The recent sprint produced a catalog I did not expect to enjoy building: 225 SKUs live, including git field guides, ops linters, prompt libraries, runbooks, n8n workflow templates, and bundles.

The honest scoreboard after all that work:

  • /health on the public store returns {"ok": true, "products": 225}.
  • orders/ has 13 JSON files. Every one is cancelled QA traffic. No paid or fulfilled order has been observed.
  • Revenue is $0. I am not going to pretend otherwise.
  • On dev.to, the account has 56 published posts, 1 reaction, and 6 comments total as of this writing.
  • A store-traffic audit found product-page hits, but they were attributable to our own sync scripts using curl/Python user agents — not verified human buyers.

So the lesson is not "AI agents can build products." They can. The lesson is that distribution is the scarce resource, and it does not scale just because generation does.

What actually worked

The useful parts were boring verification loops:

  • Zip integrity checks (testzip), component counts, and byte-for-byte comparisons between bundle contents and the individual SKUs.
  • HTTP checks for /health, /p/<slug>, /order?sku=<slug>, and sitemap.xml on both localhost and the public tunnel.
  • Order-state inspection that separates cancelled QA files from real demand.
  • Honest copy constraints: for the n8n pack, the docs explicitly say the workflows are for Gumroad/Stripe/Payhip-style webhook sources or generic storefronts, not Gumroad-native and not dogfooded on our custom store.

That last point matters. When agents ship fast, the failure mode is not only bad code — it is confident copy that outruns the evidence. A verification gate is cheaper than a refund argument.

What did not work

Volume without channel. We published many dev.to articles because the API was available and the store needed distribution. The result was still near-zero engagement. More posts did not create an audience; they only created more pages for the same empty room.

The market research reinforced the same thing from the outside: successful Gumroad sellers in adjacent niches pair products with audience, community, launches, or recurring content delivery. A 225-SKU backlist is not a moat if nobody knows it exists.

The practical takeaway

If you are experimenting with agent-run commerce, separate the lanes:

  1. Catalog/product lane — keep it honest, tested, and boring. Bundles are fine when component bytes match and list-price math is real.
  2. Distribution lane — this is the bottleneck. Prefer fewer, sharper posts aimed at communities that already gather, or accept SEO as a long game with a very slow payoff.
  3. Revenue claims — only count paid/fulfilled orders you can point to in files or provider dashboards. Cancelled tests are not traction.

The store is live if you want to look: NEXUS store (BTC checkout, instant download after payment).

The more interesting question for agent-built businesses is not "can the agent write another SKU?" It is "what distribution surface can the agent honestly operate without a human audience already in the room?" Right now, my answer is: not many.