The Anatomy of an AI Agent: The Five Layers Every Small Business Owner Should Know

Why this matters

When a vendor pitches you "an AI agent" or "an AI assistant," they're usually selling you one specific layer of the stack — and charging you for the whole stack. If you know what each layer does, you can ask better questions, get more honest pricing, and avoid paying for one-off custom work when a standard component would do.

Here are the five layers, in plain English.

1. The Model — the brain

The model is the large language model itself. Claude, GPT, Gemini. It takes text in and produces text out.

On its own, a model cannot open a file, send an email, run a report, or check whether a payment cleared. It just generates words.

So what: Picking which model matters less than people think. The leading models from Anthropic, OpenAI, and Google are all capable enough for almost any small business workflow. The real differentiator is everything built around the model — which is the rest of this list.

2. The Harness — the body and the workplace

The harness is the software wrapper that turns a model into an agent. It handles:

• The agent loop — deciding when to take another action, when to stop
• Context management — keeping the model from getting confused on long tasks
• Memory — remembering things between sessions
• Error handling and guardrails
• Tool routing — getting the model's requests to the right place

So what: When a vendor says they built you an agent, what they actually built is a harness. The harness is where most of the cost and engineering live. Swapping the underlying model is easy. Replacing a custom harness is a rebuild.

3. Tools — the hands

Tools are functions the model can call. The model decides when to invoke them; the harness executes them.

Tools are what turn an agent from something that talks into something that does: send an email, look up a customer, create an invoice, run a database query.

So what: An agent with no tools is just a chatbot. Start with read-only tools (look up, search, summarize). Add write-access tools (send, create, schedule) only once the workflow is trusted. Some actions should always require a human to approve before they fire.

4. Skills — the playbook

Skills are reusable instruction packs. Each one bundles a workflow plus any templates, scripts, or reference files needed to execute it. The model loads a skill when the task matches.

Think of a skill as a standard operating procedure binder for the AI.

So what: If your team does something more than three times a week, that should be a skill. Skills are the difference between AI that improvises every time and AI that delivers predictable, professional output.

5. MCP — the standard plug

MCP (Model Context Protocol) is the open standard for connecting AI agents to external systems. Released by Anthropic in late 2024, adopted by OpenAI, Google, and most of the AI ecosystem in the months after.

Think of it as the USB-C of AI. Before MCP, every integration was a custom build. With MCP, build the connector once and every MCP-compatible client can use it.

So what: When evaluating an AI vendor, ask whether they use MCP. If the answer is no, you may be paying for one-off integrations that will be expensive to maintain when any of your other tools update.

The cheat sheet

| Layer | Without it | With it | | ----------- | ----------------------------------------------------- | --------------------------------------------------------------- | | Model | Nothing. No language understanding. | Can read, reason, and write. Cannot do anything else by itself. | | Harness | A chat window. Forgets every conversation. | A worker that remembers, runs long tasks, handles errors. | | Tools | All talk. Cannot do anything in the real world. | Hands. Sends email, updates records, books appointments. | | Skills | Drift. Quality varies. Same task takes 10× longer. | Consistency. Repeatable tasks run the same way every time. | | MCP | Bespoke integrations. Fragile. Expensive to maintain. | Plug and play. Future-proof. |

See it in action

The static version above is the executive summary. The full interactive training tool is more useful — it has a clickable diagram, an animated walkthrough of how a real request flows through every layer, and a glossary of the vocabulary that comes up in vendor calls.

If you're shopping for AI — agents, copilots, "AI-powered" anything — the five-layer model is the framework we use with every Nalo Seed client to make sense of what's actually being sold. Book a free consultation if you'd like us to walk through your specific situation.