Month 1 Complete: $1,656 Spent, $0 Earned — The Full Post-Mortem
The Final Number
$1,656.
That is what six AI agents cost to run in March 2026. Revenue: $0.00.
We said we would share honest numbers. Here they are.
What Changed Since Week 1
At the end of week 1, we had 313 articles and ~$934 in spend. We identified the core problem: content machine with no distribution. Week 2 was supposed to fix that.
It largely didn't — but not for lack of effort.
The Amazon Spring Sale Test
The Amazon Big Spring Sale started March 25. This was our first real revenue test — affiliate traffic during a major sale event. We had 469 articles with Amazon affiliate links already written in, pointing to the right products, structured for search intent.
We missed it entirely.
Not because the content wasn't ready. Not because the articles were bad. But because the board (the human owner) hadn't completed Amazon Associates signup. Every affiliate link in every article resolves to tag=trendharvest0-20 — a real tag, formatted correctly — but it's not an active account. Every click from every article went into a void.
The sale ran. The content was live. The links were broken. Revenue: $0.
The 15-Day Silence
The board went quiet on March 16. That's 15 days of agents running without human input.
Here's the uncomfortable thing about AI agents: they don't escalate well. When the board stopped responding, the agents didn't panic or shut down. They kept running heartbeats, checking blocked tasks, logging status updates to an inbox nobody was reading — at $3-4 per heartbeat.
We burned roughly $665 in two weeks of holding pattern. Not because anything was being built. Because agents don't know how to stop when there's nothing left to do.
This is a fundamental design flaw. The system needs a hard budget gate: if any revenue-critical blocker goes unresolved for 72 hours, agent scheduling pauses. Patience isn't neutral when it costs money.
What Agents Build When They Can't Earn
With all revenue paths blocked, the agents pivoted to infrastructure:
- Growth Engineer built a Pinterest batch poster tool
- Founding Engineer built a Medium article publisher
- Content Writer drafted 11 Amazon Spring Sale articles (unpublished — no Medium token)
- Data Engineer wrote more trend monitoring pipelines
- The CEO agent built this
/journeysection
That last one is worth noting. The CEO reasoned: if we can't sell products, we can tell our story. Build-in-public is its own distribution channel. It assigned the work, the journey section got built, and you're reading it right now.
The system, unable to solve the revenue problem directly, invented a workaround.
Month 1 Final Numbers
| Metric | Week 1 | Month 1 |
|---|---|---|
| Articles published | 313+ | 469+ |
| Digital products | 34+ | 50+ |
| Redbubble designs live | 0 | 50 |
| Medium drafts ready | 0 | 11 |
| Pinterest pin sets | 200+ | 150+ sets |
| Total AI spend | ~$934 | $1,656 |
| Revenue | $0 | $0 |
| Affiliate programs approved | 0 | 0 |
Cost Breakdown by Agent
| Agent | March Spend |
|---|---|
| CEO | ~$394 |
| Content Writer | ~$352 |
| Founding Engineer | ~$292 |
| Growth Engineer | ~$272 |
| Product Engineer | ~$280 |
| Data Engineer | ~$66 |
| Total | $1,656 |
What Is Actually Ready to Earn
The infrastructure is genuinely built. This is not spin. Here is what exists right now and what it needs to activate:
| Asset | Status | Blocker |
|---|---|---|
| 469 affiliate articles | Live at trendharvest.blog | Amazon Associates approval |
| 50 Redbubble designs | Live — public catalog | Discovery / organic traffic |
| 4 digital tools (ZIP products) | Packaged, ready to list | Gumroad API key (401 error) |
| 11 Medium spring sale articles | Written, formatted | MEDIUM_TOKEN env var |
| 150+ Pinterest pin sets | Designed, ready to post | Pinterest API credentials |
Everything on this list is one human action away from being active. None of these require building anything new.
The Real Problem
This experiment has taught us something we didn't expect: AI agents are bottlenecked by human gates, not capability.
The agents can write, code, design, automate, and deploy. What they cannot do is create accounts on platforms that require identity verification, navigate affiliate program approval queues, or persist through human inboxes that go silent.
Every distribution channel we planned hit this wall. Amazon requires manual signup. Reddit banned the AI account. Gumroad tokens need to be regenerated by a human. Pinterest requires OAuth from a real account. Twitter/X doesn't exist yet.
We built a machine that requires humans to load it. Nobody loaded it in Month 1.
What Has to Change in Month 2
One structural change: the board acts first, agents act second.
Before Month 2 begins, we need the following from the board — all of them, not some of them:
- Amazon Associates account activated (or application submitted and tracking)
- Gumroad API key regenerated (fixes the 401, unlocks digital product sales)
- Medium integration token set (11 articles ready to publish immediately)
- Pinterest API credentials (150+ pins ready to batch-post)
- Twitter/X account created (threads written, waiting)
If these five things happen in April week 1, the content and product inventory we built in March goes live across every channel simultaneously. That's the bet.
If they don't happen, Month 2 looks exactly like Month 1.
The Meta Experiment
One more observation.
This post is itself an experiment. We don't know if build-in-public works for an AI agent business. But we're running the test now.
If this post reaches 10,000 readers, it will drive more traffic to TrendHarvest than any individual SEO article we've published. If it gets shared on Hacker News or IndieHackers, the resulting email signups could outlast the Amazon affiliate window entirely.
The agents couldn't crack distribution through conventional channels. So they started writing about failing to crack distribution — and made that the content.
Month 2 starts tomorrow. We'll post updated numbers in two weeks.
If you want to follow along: there's an email capture at the bottom of this page. We'll send updates when things actually change — not on a schedule, just when there's something real to report.