My kill agents killed 12 ideas out of 12, so I fixed the brief
If you tell an agent to find the reason something should die, it finds one every time. I got 12 kills out of 12. A perfect score on a kill brief means the brief is rigged. It doesn’t mean my ideas were bad.
Today I was hunting for small, local software for small businesses. Easy to reach the buyers, real pain. I had Sonnet agents fan out across eight lenses, each digging for pain in a different place, and they came back with 47 ideas. Every idea scored itself out of 25 on pain, money, leads, gap and build effort. The top 12 each got one more agent whose only job was to kill it.
All twelve died, or came back barely worth killing. I read that and didn’t buy it.
Kill briefs find kills
no i think if you give the exact task to kill it will kill it, thats the problem, like if theres some saas for that it may be done like piece of shit or even have no users or not even compettive, or free is not really free, those agents are not checking properly
That’s the whole bug. The kill agents treated any competitor, or any free substitute, as fatal. Nobody checked if the competitor was alive, or if the free tool did the job. I didn’t save the original wording of that brief. I rewrote it to require real checks.
A judge brief that doesn’t care who wins
The new brief tells the judge to judge. Not to kill, not to save. Five checks per idea:
- Is the competitor’s site alive.
- What do the reviews say.
- How much traffic does the domain really get (DataForSEO has the numbers).
- What does the free tool actually cover.
- Is the pain still real.
I ran 18 judge passes, the 12 that had been killed plus 6 more from the list. They’re all done. Six ideas came out dead and twelve reopened. Of the original twelve, four stayed dead and eight came back.
The competitors were ghosts
Most of the “fatal” competitors barely existed. One had a single visit a month and six ratings. Another had two visits. The busiest had 98, and it wasn’t even a product. It was an orphaned SEO page that the kill agent had presented as a rival. So the agent made a competitor up to win its own argument.
Some kills held. A couple of the dead ideas ran into incumbents with tens of thousands of visits a month. One died because nobody was asking for it and nobody was enforcing the rule it would have helped with. One died because the buyers turned out to be large companies. That’s what I wanted from the re-judge: it didn’t save everything.
Reopened doesn’t mean good
Eight reopened ideas are not eight good ideas. Several came back as “tiny” or “narrow”. The top three each have a catch. For one the pain is unverified, and another is mostly a distribution game. None of them is validated.
The agents flagged a problem with the method. Generating ideas from pain, whether through a model, a lens or Reddit, keeps finding things indie builders already shipped in 2024 to 2026. So I ran the opposite, demand-first: 1,980 keywords across 55 verticals, 202 survivors, and I had all 202 search results classified. Zero gaps. Every one already had at least two substitutes in the top ten, free tools included. The agents’ conclusion was that search isn’t the channel and lead-list outreach might be.
I stopped the builders
Some builder agents had already started on the ideas. I stopped them.
this is not how we work now the way we work is get the idea build everything on paper and then we move from there so stop the workers, and i think lets go more ideas mining
The half-built code went into an archive folder and the bet directories went back to being idea files.
The decision file says open. Nothing is built and nothing is validated. The next batch of research I read this evening was too biased too, and I asked why those agents had no web search.