There's a lot of talk right now about how companies will shrink down to small teams with a single AI agent running most of the business. I don't think that's quite right, not because AI can't do a lot of the work, but because of what actually happens when you centralize intelligence.

‍

The thing that makes organizations produce great work isn't that everyone agrees. It's that you have people with slightly different values and views who have to collaborate anyway. Some of those differences are about taste. Some are about risk tolerance. Some are about what the data actually means. The tension between those views is where the good stuff comes from. There is an emergent property of different perspectives with a shared goal.

‍

I think agents will work the same way.

‍

The way I see it, there are two kinds of context an agent needs:

‍

1. There's the factual stuff — what's in the database, what happened last quarter, what the product actually does. That should be shared across the whole organization. One layer, everyone gets it.
2. And then there's everything else. The judgment calls. The things where the information isn't clear enough to have a right answer, either because we don't have enough data internally or because the market just isn't legible yet.

Number 2 is what makes up the majority of decisions in something like marketing. IE which creative direction to pursue, how to read a weak signal, whether an experiment is worth the cost.

‍

For that, you need agents that have developed their own context. Not just the shared layer, but their own extended set of views built up over time through the work they've actually done. And those views should be different from agent to agent.

‍

That means you probably want multiple agents even within a single department. Each one doing part of the work, accumulating a different context graph based on their specific experience. Then the value comes from those agents contributing to each other — almost like a team of people who've each developed real expertise and have to figure out together what to do.

‍

This is what I'm thinking about with the Pearmill. Our goal is to build something that carries institutional knowledge across time and people. But I want to be careful not to build just one big shared brain where everyone converges on the same views. That would actually make it less useful.

‍

One shared layer of facts. Multiple agents building their own judgment on top of it.

‍

That's the architecture I think gets you somewhere.

‍

I’m talking about this, and more, tomorrow on a Webinar for Agentic Marketing. Would love to see you there!

‍

Agentic Marketing Webinar

‍

July 15th | 11am EST

‍

Register here

‍