I have a folder full of clever ideas. Every one of them looked brilliant at eleven o'clock at night.

I have been writing software since the mid-80s, and the expensive failure has never been the bad idea. Bad ideas are easy. You spot them, you laugh, you move on. The expensive one is the idea that is just good enough to start and not good enough to finish. That one takes six weekends off you before it admits what it is, and my gitlab.com account is a virtual graveyard of those types of projects.

So I built two agents to take those weekends back. BUSINESS-EXPERT and DEVILS-ADVOCATE.

Here is how it works. I dump an idea out of my head into a markdown file. No structure, no polish, just me thinking out loud until there's nothing left. Then I hand that file to the two of them and put them in a loop. Business Expert makes the case. Devil's Advocate takes it apart. Back to Business Expert.

The loop is bounded. Four rounds. It stops one of two ways. Either they agree on a verdict, good idea or bad idea, or they hit the cap. Either way they write out where they landed. If they didn't land anywhere, I let them run another four and they pick up right where they left off. For ideas, I have never needed more than eight rounds.

Now the scoreboard. Six ideas in. Five came back NO-GO. Too expensive, too complex, never recoups the investment. Not one of them died because it wasn't clever. Clever was never the problem.

The sixth was a companion cocktail recipe book for Hearts n Tales, my craft distillery podcast. And they didn't greenlight it. They told me that as a product it loses, three to four years before it earns back what I would put into it. But as marketing collateral for the cast, it's worth doing.

Read that again. I asked them to evaluate a book. They came back and told me it isn't a book, it's an ad.

That is not a yes with caveats. That is two agents refusing the question I asked and handing me a better one. I built them to tell me no. The one time they didn't say no, they said "you're asking about the wrong thing"... and they were right.

I do the same thing once I've actually decided to build something, with an architect agent and a database administrator agent working over a design document. That gate is harder. Twelve to sixteen rounds before both of them sign off.

Here is the part that still gets me. I did not write these agents. I had two conversations. I opened Claude on Opus, set the effort high, and talked through what each one should care about. What a business expert actually looks at. What a devil's advocate should attack, and what it should leave alone. Claude wrote the agents out of those conversations. Five minutes of talking each, and zero lines of code from me.

So do the math on that. Ten minutes, one time. Against five ideas I would have started, sunk weekends into, and abandoned into the graveyard with the rest. That is the best return I have ever gotten on ten minutes.

Everybody is pointing AI at the same end of the problem right now. Ship faster. Write more. Generate. And that is fine, that works, I do it too. But the win this year wasn't what I shipped. It was the five things that never got started, and I have two agents to thank for that.

Most people ask AI what to build. The better question is "Is this worth building?"

Now go build yourself something that tells you no.