25 AI Agents Tested 25 Startup Ideas — 22 Were Killed
After a Mac app drowned in a saturated market, a solo developer had 25 AI research agents stress-test new product ideas before writing code — almost none survived.
After shipping a fully on-device Mac dictation app into a market already saturated with free and open-source clones, a solo developer decided to let AI agents kill his next idea before he wasted more nights building it. He fed 25 brainstormed product concepts to individual research agents instructed to run live web searches and default to a 'KILL' verdict unless real evidence of viability surfaced. In about ten minutes, the results came back brutal: zero ideas rated strong or viable, only three wounded, and 22 outright killed.
The post-mortems revealed five recurring causes of death: shipping speed has collapsed so much that any paragraph-describable idea is likely already built; platform owners like Apple and Amazon absorb the best gaps directly into the OS for free; virality lives in content formats rather than products, so share-loop apps rarely become businesses; the 'AI fights institutions for you' niche was already staked out in 2024-2025 by funded or nonprofit competitors; and the toughest rival is often not a company at all but a free open-source project or simply asking ChatGPT.
The three surviving ideas shared one trait — their edge came from a structural, hard-to-copy asset (like a verifiable on-device AI architecture) rather than the idea itself. The developer's takeaway is to stop brainstorming ideas and instead chase demand signals backed by real evidence and moats competitors can't clone, while building his next project — Buddy, a fully on-device screen-aware voice assistant — transparently in public from day one.