What agents actually are.

If you've encountered "agents" through a Microsoft Copilot demo or a LinkedIn carousel, you've been shown a wrapped-up box. It does some impressive things. The contents are intentionally a little blurry.

This page opens the box. Each layer above is a working demo. By the end you'll know why some agent products feel like magic and others feel like a glorified search bar.

The difference is the harness around it.

A large language model is a text generator. You give it text. It gives back text. That's the whole contract.

The model has no access to your purchase orders, your suppliers, your work orders, or the internet. It has no memory and no way to act. It's a function from text to text.

Run the demo below and notice what happens, and what can't.

The model tried to answer. It might have admitted it doesn't have access to your systems. It might have made up plausible details. Both are correct behaviors for a text generator.

The part most explanations skip: anything an "agent" actually does for you (looking up data, calling APIs, taking actions) is being done by something else, not by the model. The model generates text that describes what should happen. Something else reads that text and does the work.

The fix is a loop. Give the model a list of tools it can call. When it generates text that looks like a tool call, parse it, run the tool, feed the result back. Repeat until the model produces a final answer.

This is what most products you've used call "an agent." Most chatbots-with-tools and "Copilot"-style features are this. It's an improvement. The agent can now look things up and take simple actions.

It's also where most products stop.

The model decided what tool to call, the harness ran it, the model got the result, and so on. The model is doing more, but still the same kind of thing: generating text. The harness is doing the work, and the harness in this layer is thin: a tool registry and a loop.

This is the layer where most "agent" claims live today. It works. It's also the ceiling for most products you've encountered.

The light harness has a real limitation: it doesn't know what your business is. Every conversation starts from zero. Every tool call re-fetches context. The model reasons about your operations from whatever fits in one message.

A real harness keeps a graph of how your business works: purchase orders, suppliers, parts, work orders, customers, and the relationships between them. We call those dimensions. When a question arrives, the harness traverses the relevant dimensions, builds a focused context, and runs the model across multiple steps that share state.

The simulation below runs the same PO #12345 question through a small version of this harness. Five nodes (PO, supplier, work order, part, alternate supplier), real traversal. Watch the harness investigate.

The harness didn't just call a single get_po tool. It walked from the PO to the supplier, looked at recent on-time history, walked to the work order, checked the deadline, walked to the part, found an alternate supplier, and composed a recommendation with a concrete next action.

The model still only outputs text. The harness is doing all the work, butmore interesting work: graph traversal, context construction, multi-step coordination. And it's working with a model of your business, not just a list of API endpoints.

That's the difference between layer 2 and layer 3.

Most products you've seen called "agents" are layer 2. They're a model with a tool loop wrapped around it. That's a real technology and you should keep using it.

When someone tells you "we built an AI agent for [your business problem]," ask which layer. Layer 2 with a custom prompt is fast to build and limited. Layer 3, a harness with a graph of how your business works, is what you need when the question crosses three or four parts of the operation at once.

myai is layer 3. The simulation above is a tiny version of what the platform does. If you want to see it with a graph of your actual business (your POs, your suppliers, your customers, your work orders), book a conversation.