agentmunger: a new way to look for undervalued opportunities
Charlie Munger said that if you mix raisins with turds, they’re still turds. He was talking about diversification, but the line does more work than that. It is a warning about averages. Put one good thing in a basket of bad ones and the basket does not get good — you have just made the bad harder to see.
That is what most screening does. You sort by a ratio, you get a list, and the list is mostly turds with a few raisins in it. The ratio was never the insight. It was a filter you could compute cheaply, and cheap filters find cheap things. Cheap and undervalued are not the same word.
Cheap is a number, undervalued is a judgment
An asset is cheap when a number is low. An asset is undervalued when the market has the story wrong — when the thing people believe about it is not what is true. The first is arithmetic. The second takes reading: filings, transcripts, court records, hiring pages, the paragraph of a 10-K that nobody got to.
That reading is where the work has always been. It is also exactly the work that did not scale. One analyst covers a handful of names properly, or a hundred names badly, and the economics of the business push hard toward a hundred badly. So the screens got better and the reading got worse, and the industry agreed to call the screens research.
What we’re building
agentmunger is an attempt at the other half. Not a faster screen — there are plenty, and they are nearly free. Agents that do the reading: that go through the primary documents, hold a thesis in mind, and come back with what actually contradicts it.
Three things we care about more than coverage:
Disconfirmation over confirmation. It is easy to build something that finds support for whatever you already think. Harder, and more useful, is something that goes looking for the fact that kills the idea. A thesis that survives an honest attempt to break it is worth something. One that was never attacked is not.
Sources you can click. Every claim traces back to the document it came from. If a judgment cannot be checked, it is not research, it is atmosphere.
Saying nothing, often. Most names are not mispriced. A system that always has an answer is telling you about its incentives, not about the market. We would rather return an empty list than a confident one that is wrong.
Where this goes
The honest version of the pitch is that this is hard and we are early. Reading is not a solved problem, and anyone claiming otherwise is selling a screen with better manners.
But the shape of the thing is clear enough. The constraint on real research was never ideas, it was attention — how many documents a careful reader could get through before the quarter ended. That constraint is moving for the first time in a long while.
What it does not change is the standard. Mix raisins with turds and you still have turds. More output is not the goal. Fewer, better-defended ideas are.
Join the waitlist if you want to hear when we open it up.