It always comes back to network effects
Agents are going to raid all of your data.
Systems of record are no longer defensible. For decades, the information inside data-heavy apps was trapped, because moving it cost more than just living with the system you had.
But now, sufficiently capable agents can just raid the systems of record by manually extracting all of the data, row by row, and putting it elsewhere. It might take time, but agents have unlimited time (as long as their human is willing to keep paying for their tokens).
A system of record is really just a database of facts, and facts are portable. There are AI startups that are actively deploying agents to systematically tear down a customer’s CRM on another platform. It’s kind of brilliant. And this playbook will repeat for other types of applications, too: chatbots’ memories, agent harnesses’ customizations.
Network effects are the answer. Like they’ve always been. To make a system of record actually defensible in an agent-first world, you have to have network effects.
An agent can copy every row in your CRM, but it can’t make your customers open the new vendor’s emails.
An agent can scrape every post, photo, and follower graph off Instagram. But it cannot make your friends open a different app to see what you posted.
An agent can distill a model. In fact foundation models themselves are becoming commoditized. Every lab is converging on similar capabilities. Even the data a model was trained on is portable (the open-source community has proven this repeatedly). What isn’t portable is the user base actively using a product and generating the preference and interaction data that makes the next version better.
It always comes back to network effects.


Networks cannot be built by agents, so strong agree at the high level.
Similar network binary thesis to Travis (you either have network effects and survive or you don't and die): https://travismay.medium.com/moats-matter-again-a6359b0d31da