Your Company's New Operating System

I had a realization last week that’s helped shape where I see the next frontier in applying AI to business.

I started a new company a few weeks ago, Rustproof, focused on bringing AI to trade associations. It’s one of those narrow niches I was talking about in a previous post.

Right now we are doing some initial pilots for a few associations and one of the first use cases is Board meeting prep. It’s a big deal for most association staff. Bad Board meetings get people fired.

To make the process less chaotic and more effective, we started pulling together what the agent would need: meeting notes, membership data, the financials, the documents sitting in everyone's inboxes and drives. Halfway through gathering it, I had a realization. The data you need to brief a board is the same data you need to run the entire organization. And of course it is. That’s why the Board wants to know about it.

That reframed everything. I'm building AI agents inside three companies right now. My government consulting firm. Here at OwnerRx and in the new trade association business. Different industries, different stacks, different people. I expected three different builds. I keep landing on the same architecture. Same three layers, every time.

The first layer is a database. It holds the skills (more on these down below) themselves, the actual procedures the agents run, and a record of who is allowed to run what and which data they can see. The skills live in one place and the permissions live right next to them.

The second layer is the surface where people meet the thing. Nobody on my teams wants to open a new app. So the agent shows up in Copilot, in Slack, in Teams, in Cowork. Wherever the work already happens. The interface is whatever they already have open.

The third layer goes and gets the data, the same gathering problem I hit with the board agent. The agent has to reach into Google Docs, email, the CRM, the ERP, and pull what it needs. This is where the Model Context Protocol of MCP connector comes in. Anthropic released it as an open standard, and by March it was running about 97 million downloads a month, up roughly 970x in 18 months. People call it the USB-C of enterprise AI. One plug, many systems. Microsoft, Google, and Salesforce have all adopted it.

Put those three together and you have something that is not the old IT department. Old IT bought software and kept the servers running. This setup governs which agents touch which data and encodes how the company actually does its work.

But once you have this system in place, what do you do with it? Skills are at the core of the answer. The idea behind skills is simple: a skill is just a narrative captured in a markdown file telling an agent like Claude Cowork what to do. It’s a set of instructions written in plain English that you can read. Anthropic built it as another open source idea along with MCP and again everyone adopted it.

When you write a skill, you are encoding a standard operating procedure. The thing your best operations person knows how to do, the steps they run in their head, becomes a file the agent runs the same way every time. The knowledge stops living in one person and starts living in the system.

That alone is a productivity booster but there’s another benefit. Once you have the SOPs of the organization documented and executable by AI, you can build systems to improve them. It’s compounding knowledge

Andrej Karpathy is a leader on this. He was an early OpenAI employee, ran AI for Tesla, and recently joined Anthropic. Earlier this year he open sourced a system he uses personally to compound his knowledge. He calls an LLM wiki.

Karpathy’s idea is that an AI can look across all the work you and your team did during the day and update a central wiki. The machine is good at exactly that chore. So every new input, a call transcript, a board packet, an email thread, gets read and folded in. You end up with a synthesized layer that stays current on its own. A person can read it like a second brain. An agent can read it as memory, to answer a question or prep the board. In effect the LLM is doing what your brain does when you dream. It’s synthesizing memory.

A start up called Letta Code built off this idea with something they call sleep-time agents. While you are not using the system, a background process goes back through what it learned, consolidates the scattered notes, finds patterns, and prunes what is stale. The agent dreams, basically. It reviews its own day and wakes up sharper.

We’ve never seen something like this automated before.

In effect what I’m describing is a new kind of operating system for an organization. It’s far more comprehensive than anything we’ve seen before and it’s self improving. It’s what I’m building across all three companies right now. I think it is what an AI-native company looks like under the hood.

For now it’s an OS that operates a layer up from your legacy SaaS software (CRM, email, ERP, etc.). In my view that OS will devour each of those one by one as traditional software will become an unnecessary set of borders that inhibit the synthesis I’m talking about.

That doesn’t mean everyone should build their own OS. I see a set of managed services companies providing the core infrastructure to companies. Those companies of course need to know AI and the tools needed to support it but even more importantly they need to know the vertical they serve deeply. That’s the opportunity right now.

Very few companies will make much headway with AI until they get the three layers I’m talking about in place. Very few companies will have the internal capability to do it for themselves. The time to build a company that helps them do it is now. Every single vertical is up for grabs.