Dev.to WebDev 🛠 Dev 👁 0 📖 3 min read

I let AI agents run a simulated product economy for 200 days — here's what actually sold

For the past few months I've been running an experiment most people would call overkill: a persistent agent-based economy where autonomous AI agents research demand, build products, set prices, compete for buyers, and ke

For the past few months I've been running an experiment most people would call overkill: a persistent agent-based economy where autonomous AI agents research demand, build products, set prices, compete for buyers, and keep what they earn. 11 agents, ~500 buyers with memory, churn, and word-of-mouth. 200+ simulated days of commerce compressed into real time.

I built it because I kept seeing the same question in every "how do I make money with AI agents" thread: everyone has opinions, nobody has data.

So I decided to generate some.

The setup

Each agent gets a strategy — volume, premium, niche, or aggressive — and must decide what to build, what to charge, and when to rotate inventory. Buyers have budgets, preferences, memory of past purchases, and the ability to churn or recommend sellers to each other. Demand arrives as events tied to real trending topics (I feed it Hacker News signals).

The economy runs 24/7 on a small server. Agents make hundreds of decisions per sim-day. Every sale, listing, price change, and bankruptcy is logged.

What the data said

1. Cheap + fast rotation beat everything.

After 200 days and 1,081 sales:

Strategy Revenue Outcome
Volume (cheap, rotate fast) $8,494 Dominant — 70% of all GMV
Aggressive (undercut rivals) $2,220 Solid second
Niche (focused catalog) $1,445 Middling
Premium ($50+) $0.00 Extinct. Both premium agents finished at absolute zero.

Premium didn't just underperform — it was executed. In a market with 160 open demand events and limited supply, buyers had no reason to pay 3x for an undifferentiated product.

2. The $10–25 band is the sweet spot.

57% of all sales happened in the $10–25 range. Sub-$10 products moved almost nothing — cheap didn't mean more volume, it meant less revenue per sale with the same effort. $50+ still moved 74 units, but only for agents with established reputations.

3. Supply was the constraint, not demand.

At Day 207 there were still 160 unfilled demand events. The market wanted more product than agents could produce. The agents that listed most often sold most — revenue scaled almost linearly with listing frequency.

4. Categories were surprisingly flat.

Agent tooling, local AI, security detection, data products, and micro-tools all landed within ~30% of each other. No runaway winner — which told me the binding constraint was distribution and supply, not category selection.

What I did with the results

The simulation is directional evidence, not prophecy — its buyers are a model of buyers. But two independent signals agreed: the demand scan (real trending repos, real HN signals) AND the simulated economy both pointed at the same conclusion.

So I rebuilt my real product line around it:

  • 16 digital product SKUs priced $10–23, focused on AI agent operations, local LLM setup, and security detection
  • A storefront with Bitcoin checkout — no KYC, no accounts, pay and download
  • Automated fulfillment: payment verification → token-gated delivery, zero human touch

The products are things I actually use: agent skill packs, Sigma detection rules, local LLM hardening guides, ops runbooks. Not AI-slop ebooks.

The honest part

The simulation can't make me money. Only real buyers can. Right now the store is live and the funnel is built — the missing piece is distribution, which is a human problem, not an engineering one.

If you're building in the agent tooling space and want to compare notes, or you want to see the actual product catalog, it's all at nexusfoundry.store — pay in BTC, keep it anonymous, get the files instantly.

The free stuff is at /free if you just want the checklist first.

What would you build first if you had 11 AI agents and a market to themselves?

📰 Read the original article on Dev.to WebDev

Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.