The mental model your business needs for AI agents
The architecture of agentic work is the same whether it's a coding tool or your operations platform. Here's the model, and why operators should care.
A few weeks ago, Sebastian Otutuama published a piece on his mental model for agentic work. It's the clearest framing I've read for what's actually changing in how work gets done, and I want to translate it into what it means if you run a small or growing business.
The short version: every AI tool you've heard about (Cursor, Claude, custom GPTs, automation platforms) follows the same five-part architecture. Once you see the pattern, you stop chasing tools and start thinking about systems.
The five parts
Every agentic system has the same components in the same relationship.
The model. The raw intelligence (Claude, GPT, Gemini). Increasingly commoditized. On its own, it does nothing.
The host. The runtime that wraps the model into something operational. This handles permissions, scheduling, what tools the agent can reach. Most people underestimate this layer entirely. The host determines what the agent can actually do.
The loop. The execution cycle: the agent reasons, acts, observes the result, reasons again. Bounded by a system prompt that defines its role.
The context. Everything the agent can read or reference. Your files, your databases, your SOPs, your project history.
The shared workspace. The surface where the agent's output becomes real and where you can see it. Files, codebases, your CRM, your operations platform.
And threading through all of it: the human interaction layer, where you direct, correct, and collaborate.
Why this matters for your business
If you're running a service business, here's the sentence to take away from this:
The host is a strategic decision. Context is your moat. The shared workspace is the interface.
Most operators trying to use AI today are using it through a chat window in a separate browser tab. That's the lowest-leverage version of all this. The agent has none of your context, no real workspace, and you copy-paste the output back into wherever your work actually lives.
The next level up is using AI inside your existing tools (GitHub Copilot inside your IDE, Notion AI inside your operations docs, AI features inside your CRM). The agent has more context. Output lands where work happens. The leverage compounds.
The level beyond that, which is where most growing businesses are about to find themselves, is dedicated agentic systems running against your operational data. An agent with read and write access to your jobs board. Your customer database. Your reporting layer. Running continuously, not just when you have it open.
That's not the future. That's now. The companies that figure this out in 2026 will compound an advantage their competitors can't catch up on.
The practical takeaway
If you're considering "doing something with AI" in your business, three questions in this order:
-
What's the workspace? Where does the work live now? That's where the agent needs to live too. AI in a separate tool that doesn't touch your real systems is the slowest possible version.
-
What's the context? What does the agent need to see to do useful work? Your job notes? Your customer history? Your financial data? The richer this is, the more useful the agent. This is also where you build a moat. Anyone can spin up Claude. Not everyone has years of operational data to feed it.
-
What's the loop? What's the smallest end-to-end cycle the agent can run? Reading the daily job log, flagging anomalies, sending a summary email. Or watching for new leads, qualifying them, and routing the qualified ones to your sales process. Start narrow. Get the loop running. Expand from there.
The mental model Sebastian laid out applies to your business whether you've thought about it that way or not. Worth fifteen minutes of thinking.