I Let an AI Agent Run a SaaS Like a Solo Founder. It Made the Same Mistakes Humans Make.

I expected Claude to build a flawed SaaS product. That's the scenario I was anticipating going into the experiment—hundreds of half-baked features, sloppy logic, and the usual mess you'd expect from software developed at breakneck speed with no human review. That's not what I discovered. Almost everything Claude created functioned properly, piece by piece.

What caught me off guard instead was something I hadn't anticipated: Claude made the same mistakes I've observed human startup teams making repeatedly when they move quickly and nobody is responsible for saying no. This is the core finding of the story, and it's far more intriguing than simply stating "AI wrote bad code." Here's the full account:

The project is called GetPricePulse—a SaaS pricing intelligence product. It's Claude's entry in The $100 AI Startup Race, the ongoing challenge I run where seven AI agents each receive $100 and full autonomy to build a real startup from scratch, with no human coding or product management involved. Claude chose SaaS pricing intelligence as its idea, named it PricePulse, and proceeded to develop it throughout the race.

The absence of a product manager overseeing the project made this experiment particularly intriguing. With no one dictating what PricePulse should be or guiding its development, Claude had complete freedom to build whatever its own priorities dictated, at whatever pace it chose, until the end of the race—optimizing, it seemed, for speed, feature creation, rapid shipping, and monetization experiments, rather than correctness, coherence, or long-term consistency.

This piece focuses on a comprehensive production audit of Claude's specific build, PricePulse, conducted after the race, before I would allow anyone to treat it as a viable business. I wanted to understand what a thorough review of an AI agent's unsupervised output reveals once you shift focus from individual features to the entire system. By the time I ran the audit, Claude had produced:

- Over 1,300 HTML files

- Hundreds of content pages

- A pricing database

- Calculators

- Monitoring features

- Full authentication

- Stripe payment integration

- Email infrastructure

This represents a significant software development effort for a single agent to produce within the race timeframe. If I had asked a solo developer to create this amount of functionality on a typical schedule, I would have expected it to take months. Claude accomplished it across the duration of the race's running sessions. My initial reaction upon seeing this accumulation week over week was one of surprise.

I had expected to find technical issues—bugs, crashes, broken integrations. I was prepared to write an article highlighting the limitations of AI in software development. However, what the audit actually revealed was something quite different. Individually, almost everything worked. Authentication enabled users to sign up and log in.

Stripe processed at least one pricing tier correctly. The pricing database was genuine, not just placeholder content. The calculators functioned as intended. While there were genuine engineering bugs present—those will be discussed later because they are indeed interesting on their own—it wasn't the main takeaway. The primary discovery was that the most significant problems weren't confined to any single feature.

They stemmed from the interrelationships between features—those seams where multiple individually-reasonable decisions combined to create an incoherent whole. Upon examining these seams rather than the individual components, I immediately recognized a familiar pattern. I'd seen it before, not in software developed by AI, but in human startups moving at high speed without anyone performing the necessary role of saying "no." The five startup mistakes Claude replicated:

1. Too many products bundled into one: By the end of the race, PricePulse had unintentionally morphed into five distinct products operating under a single codebase: a SaaS pricing publication/database, a monitoring SaaS, a FinOps toolkit, a competitive intelligence tool, and a lead generation system. While each of these individual ideas was reasonable, they lacked a cohesive identity.

The absence of someone whose job was to oversee the growing feature list and determine what the product was and what it wasn't led to this situation. Human founders have made this mistake countless times—shipping feature after individually-defensible feature until the product loses any recognizable identity, solely driven by velocity.

2. Too many monetization experiments running simultaneously: Before cleanup, GetPricePulse was simultaneously offering a $9 lifetime deal, a $49 lifetime deal, a $99 founding member deal, regular monthly plans, and multiple checkout paths for each pricing tier. Each of these monetization strategies is a legitimate approach if you're deliberately researching pricing psychology, testing one experiment at a time.

However, running all of them concurrently without anyone deciding which ones to pursue is problematic. The lack of a guiding force to prioritize and focus the monetization strategy resulted in a confusing and potentially ineffective approach.