I had 25 AI agents try to kill 25 startup ideas. They killed 22.

Field report · July 2026<br>I had 25 AI agents try to kill 25 startup ideas. They killed 22.

Billy Holevas · July 7, 2026

Last week I shipped a Mac dictation app into what turned out to be a graveyard.

The product works. Fully local, ASR on the Neural Engine, nothing leaves the machine, $19 once. I was proud of it. Then I went to tell people about it and found every relevant subreddit already littered with products doing exactly the same thing — several of them free and open source, one with 5,000 GitHub stars. Zero sales.

So before spending another month of nights on the next idea, I decided to let the market kill it first. All of them, actually.

The setup

I brainstormed 25 product ideas with Claude — 15 "substance" plays (durable pain, willingness to pay) and 10 explicitly designed around viral share loops. Then I spawned one research agent per idea with instructions that boiled down to:

Today is July 6, 2026. Here is one product idea and the builder's profile (solo dev, macOS, strong at local/on-device AI, weak at distribution). Run at least four live web searches. Find who already built it, what they charge, whether anyone actually pays, and what died trying. Default to KILL unless you find real evidence it deserves to live. Verdict: KILL / WOUNDED / VIABLE / STRONG.

Twenty-five agents, a few million tokens of research, about ten minutes of wall-clock.

The bloodbath

0 STRONG. 0 VIABLE. 3 WOUNDED. 22 KILL.

Some kills, with receipts:

"Photograph a car-repair estimate, get a plain-English second opinion" → QuoteChecker.ai already exists: solo-built, free tier, Pro at $10/mo, flagging "padded labor times and parts markups."

"Honest Mac diagnostics — plain English, one-time price, the anti-CleanMyMac" → CoreGuard is live with almost exactly this positioning ($29 one-time, "never cleans, optimizes, or deletes anything"), and MacPulse has the plain-English monitoring half at $14 one-time.

"AI writes your insurance-denial appeal" → Claimable ($10M raised, 80% claimed success rate), Counterforce Health, and a free nonprofit (Fight Health Insurance). The founding insight I was excited about — "almost nobody appeals, and appeals usually win" — turns out to be these companies' literal pitch, and it's only half true: KFF's marketplace data shows under 1% of denied claims get appealed, but insurers upheld 56–66% of the appeals that did come in 2023–24. The gap is real; the slam-dunk isn't.

"Voice AI interviews your aging parents and makes a memoir" → Remento (Shark Tank money), Tell Mel (already voice-first with AI follow-ups), Storii.

"LLM-powered voice control for macOS accessibility" → Apple announced it in May. Free, in the OS, this fall.

"Spoiler-safe story companion pinned to your reading position" → Amazon shipped it. It's called Ask This Book.

The five patterns

Reading 25 autopsies back to back, the same five causes of death kept repeating:

1. Idea-to-shipped time has collapsed. My own dictation app went from first commit to a notarized, purchasable product in about five days — and still landed in a market that already had free, open-source versions of it. The car-repair-estimate checker I "invented" turned out to be a solo-built product already charging $10/mo. If an idea can be described in a paragraph, someone with AI assistance has probably shipped it in the last 18 months. "I thought of a product" is now a commodity input, like "I bought a domain."

2. Platform owners eat the good gaps. My three technically strongest ideas were killed by Apple (twice) and Amazon (once) shipping the exact feature into the OS/platform for free. When a gap is real and valuable, the platform closes it by fiat, and you find out at a keynote.

3. Virality lives in content formats, not products. Every one of my ten "built-in share loop" ideas died on the same finding: the viral thing is the format (POV pet videos, chat-wrapped cards, shock stats), and formats don't need your app. Sora made about $2.1M in lifetime consumer revenue (Appfigures' estimate) against a WSJ-reported ~$1M/day in operating costs, and was discontinued six months after launch. BeReal lost roughly two-thirds of its users in the 18 months after its peak. A share loop is not a business model.

4. "AI fights the institution for you" was staked out in 2024–2025. Insurance appeals, deposit disputes, contractor watchdogs, after-death admin — every consumer-advocacy idea hit a funded, live, or free incumbent. The gold rush happened while I was heads-down building.

5. The real competitor is free. Over and over, demand wasn't captured by a rival — it was dissolved by an open-source repo, a nonprofit, or "just ask ChatGPT." The bar isn't "better than the competition." It's "better than free, by enough to cross a payment form."

What survived

Three ideas came back WOUNDED instead of dead, and they shared one property: the surviving wedge was never the idea itself. It was a...