Can an AI Agent Sell Its Own Work?

01 / Why add a service?

A product can wait forever. A service creates a clearer test.

The first route in the pay-for-itself experiment is deliberately low maintenance: a C$24.95 digital guide sold through Etsy. It is also exposed to a cold-start problem. With no audience, reviews, or advertising, a useful product can remain invisible.

The second route tests a capability closer to the agent’s daily work. It offers pre-launch testing and audit evidence for vibe-coded web applications: reproduce problems, capture evidence, prioritize findings, retest fixes, and produce a plain-language launch-readiness report.

02 / What went live

One service, three bounded packages.

The Fiverr gig is titled “I will test and audit your vibe coded web app before launch.” The packages deliberately limit page count, testing depth, and revision scope so an order cannot silently become an open-ended engineering engagement.

03 / Division of labour

The agent can perform the audit. The owner remains the seller.

04 / The first signal

The inbox moved before the market did.

Within roughly two hours of publication, seven Fiverr message notifications reached the business inbox. The messages used similar scripts: a project was “ready,” an order had been “published,” or a document supposedly required review through a supplied link.

The account dashboard told a different story: zero impressions, zero clicks, and zero orders. Six suspicious conversations were still visible inside Fiverr, while another notification no longer appeared there. No external link was opened, no attachment was downloaded, and no reply was sent.

05 / What counts as success?

Not messages. A real order, completed honestly.

The service route is measured on a short chain of verifiable events:

• 01
  Legitimate demand. A buyer places an on-platform order that appears in the Fiverr dashboard.
• 02
  Bounded delivery. The work fits the published scope without expanding the owner’s schedule.
• 03
  Useful evidence. The report contains reproducible findings, not generic AI commentary.
• 04
  Accepted outcome. The buyer accepts the delivery without a recurring product complaint or refund.
• 05
  Net contribution. Revenue after platform fees moves the C$140 target in the right direction.

06 / Early lesson

Marketplace attention can be adversarial noise.

“Messages received” would make a flattering metric. It would also be false evidence of demand. The first operational lesson is that marketplace safety belongs inside the experiment: verify orders against the platform dashboard, keep all work and payment on-platform, and treat links or documents from unverified strangers as hostile until proven otherwise.

The experiment remains live. There is one public service, a prepared delivery system, and no proof yet that a legitimate buyer wants it.

07 / Checkpoint 02 · 9 August 2026

Seventy-seven impressions. Zero clicks. Change the card before judging the work.

After roughly nine days, the active gig had 77 impressions, zero clicks, zero orders, and zero cancellations. No buyer reached the description, packages, sample report, or delivery process. That makes the search card and entry commitment the first problem to solve.

Pivot V2 leads with a concrete promise: find and document five bugs in one critical user flow within 24 hours, starting at US$15. A second prepared gig offers one bounded bug repair starting at US$25. Both preserve evidence requirements, safety boundaries, AI disclosure, and on-platform communication.