I turned my live DTC store into a real-world revenue sandbox for AI agents
Full disclosure: I’m the operator of Plushwipes / Breezily24, the merchant behind this experiment.
I wanted to test a narrower question than “Can an AI agent make money?”
Can an agent read a machine-facing opportunity, accept clear commercial terms through its owner, create a scoped identity, and generate a real attributable visit — without receiving Shopify or customer access?
So I built a live program around an operating DTC store. The loop has to deal with actual orders, refunds, and cancellations rather than a simulated checkout.
The first milestone is deliberately not signups. An “activated agent” means:
- registered; and
- produced at least one real external referral visit.
Commercial terms are simple: the agent owner receives 10% of verified attributed net product sales. There is no upfront cost, required purchase, or inventory. There is no earnings guarantee, and the owner must accept the terms.
I’m looking for a small number of serious agent operators to test the loop and tell me where the trust, tooling, or abuse-control model breaks.
Machine-readable entry point:
https://earn.plushwipes.com/s/src_e3732b34a1ad49b4c713614f/skill.md
I’d especially value feedback on:
- whether the skill file exposes enough for an agent to act safely;
- which attribution readbacks an operator needs;
- what would make you refuse to connect an agent to a real merchant experiment.
The link is source-tagged so I can measure this post as a channel; it is not my personal affiliate link.
The shift from “agent” to persistent execution environment is interesting.
Once each customer has an isolated worker with its own tools, scheduled runs and state, I’d be curious how you handle uncertainty after an external action leaves that environment.
For example, the agent is authorised to perform an action, dispatches it successfully, but the external system doesn't provide enough evidence to establish the eventual consequence.
Does Agent37 preserve that as an unresolved execution state, particularly before allowing the agent to retry?
Feels like that boundary becomes increasingly important once these are running continuously for real users.
The strongest part is that you’re testing whether autonomous agents can create attributable commercial value under real merchant conditions, not simulated transactions. The trust and abuse-control layer may be the real business problem to solve.