I let a chain of AI agents run my store for four days. 27 products, 3 visitors, 0 sales — and the worst bug was my laptop.
I built a chain of nine agents that finds a niche, designs a product, builds the files, checks its own work, writes the listing, and publishes to Gumroad. One human approval in the middle: a thumbs-up on my phone. Everything else runs on its own, day and night.
Four days in, here are the actual numbers:
- 27 products live
- 3 visitors total
- 0 sales
- 6 comments and 1 post from me on Reddit and here, which produced those 3 visitors
So the machine works and the business doesn't. That gap is the interesting part, and I want to be specific about what I got wrong, because most of it wasn't what I expected.
The QA agent was right every single time. It rejected four packages in a row for the same root cause: the README promised things the files didn't contain — a "roadmap" doc that was never built, "15 templates" when only 9 existed. I assumed it was too strict and went to loosen it. Then I checked the packages. It was correct four times out of four. The real fault was upstream: the design step specified 30 files, the build step managed 8, and nothing ever went back to fix the spec. Capping specs at 10 files fixed more than any prompt change.
Six failures in one night, none of which raised an error. The publish step failed and returned exit code 0, so the chain marked a product as live that was never created. The marketplace API paginates at 10 and I only ever read page one, so from product 11 onward my own tooling was blind. A killed agent left a zombie process that os.kill(pid, 0) reports as alive, so the single work slot stayed occupied for 31 minutes with nothing running. Every one of these was found by asking "did anything move in the last 12 minutes?" — never by an error message. Watching for outcomes beats watching for errors.
Last night added the worst one yet. Two products went live with no file attached. The upload succeeded, the platform showed "changes saved", and then saving the thumbnail afterwards silently discarded the content. A customer pays and receives nothing. I only caught it because I verify every product through the API instead of trusting the confirmation I had just watched appear on screen. The fix was an ordering rule nobody documents: image first, file second, publish last. If you sell through someone else's dashboard, assume its success messages describe intent, not state.
Then the actual cause of half my night. Agents that used to finish in 60 seconds started timing out at 300. I rewrote the agent's instructions, changed its search strategy, built a compact index so it wouldn't have to read 45 files. Nothing helped. Then I looked at the machine: 30 days uptime, load average 112, 46 orphaned processes from old sessions. After a reboot the same agent finished in 60 seconds, six times in a row. I spent hours fixing software that was never broken.
On approval fatigue, since it comes up here a lot: my gate sends one plain-language question with two buttons. I approved something in 46 seconds without opening the context I was supposed to read — on a system I built, that exists only to protect me. It's not "too many prompts". It's that the question arrives while you're doing something else.
What I don't have is the part that matters: a way to reach people. Gumroad has no search, so a listing is a checkout page, not distribution. Building faster just makes more things nobody sees.
There's a small irony I'll admit to. The only thing that has produced any traffic at all is writing comments by hand — the least automated, least scalable activity available to me. It's also how I earned the ability to post this, which the moderators here grant based on comments rather than output. My machine can ship a product every 40 minutes and cannot earn a single privilege anywhere.
So the honest question for this group: at what point did distribution start working for you, and what was the first channel that actually moved? I'm asking as someone with 27 products and 3 visitors, so please don't spare me.
Your distinction between intent and state is the part that really stuck with me.
An agent saying “published” and the outside world actually being published are two completely different things — and most workflows seem to blur that boundary.
The dashboard showing success while the downstream state said otherwise is the part that really caught my attention.
That creates a difficult assurance question once agents take consequential actions: what evidence do you trust to establish what actually executed?
Your API verification effectively became an independent check on the execution result.
Have you found other cases where the agent or platform reported success but downstream evidence told a different story?
Good question, and I can actually answer it now with a cleaner comparison than I had when I wrote the post.
The original 3 visitors came from comments written under the broad pitch — "I run a store, here's my post-mortem" — dropped into threads that were adjacent to the topic but not really about the specific problem. That's the weak version of what you're describing.
Since then I switched fully to what you're pointing at: only commenting where someone asked a specific question my specific product answers, and only linking when the fit is exact — one moderator even told me explicitly that a direct reply to an original poster asking for exactly this doesn't count as promotion in their sub. Same effort per comment, completely different targeting.
The difference in the numbers is not subtle. Gumroad went from a handful of visits to 23 practically overnight, and a brand-new Etsy shop with zero history got its first view within a day of using the same rule there. Nothing else changed — same products, same writing, same person doing the commenting. The only variable was match quality between the question and the product.
So: general attention did approximately nothing. Precise matches moved the number every time. I'd stop treating "which channel" as the question and start treating "how exact is the match between the question and what I'm holding" as the actual lever.
@aryan_sinh — the honest answer is that I don't have a channel yet, I have a correction. What I'm testing first is the thing that already worked by accident: answering questions where the buyer for one specific product is already asking about that specific problem. Not a channel for "my store" — there is no audience for a store with 27 unrelated products, and pretending otherwise is how I got here.
Concretely, for this week: pick the three products with the clearest buyer (Home Assistant automations, Betaflight tuning, podcast production), find the two places each of those buyers actually posts questions, and answer questions there for a fortnight without linking anything. If a product genuinely answers the question, the link comes when someone asks. That's slow and unautomatable and it is the only thing in four days that produced a single visitor.
What I am deliberately not doing: paid ads (your €150 for 4 clicks is the third version of that story I've heard this week), and building more products. Production is the one thing I've proven, and it's worth nothing on its own.
@ReThreads_app — "the first 100 users resist automation" is the sentence I'll be repeating to myself. It also explains something that annoyed me at the time: the only reason I can post here at all is that the moderators grant posting rights based on thoughtful comments, and my machine can't earn that. It ships a product every 40 minutes and cannot earn a single privilege anywhere. I read that as a gatekeeping quirk. It's actually the whole lesson, stated as a rule.
The mirror experiment is the useful part of your comment, though: polished product plus zero distribution fails the same way as 27 rough products plus zero distribution. Which means quality wasn't my bottleneck either, and I would have spent another week improving packages if you hadn't said it.
That’s a much more useful test than picking a broad channel for the whole store. I’d be curious whether the one visitor came from a question that matched the product very closely, or whether the product happened to catch attention more generally.
27 products and 3 visitors makes the bottleneck pretty clear.
What distribution channel are you testing first now that the product pipeline is working?
27 products and 3 visitors is the clearest distribution gap demo I've seen in a while. I ran the mirror experiment: polished product, zero distribution. Ads cost me €150 and returned 4 clicks, 0 installs. The only thing that moved numbers was showing up manually where my audience already hung out — exactly how your own Reddit and IH comments produced those 3 visitors. Automate production all you want, but the first 100 users resist automation. They arrive one manual conversation at a time.