Agents 101
What agents actually are.
The model generates. The harness gives it context, tools, and a way to act.
One question. Three layers. Run each to see what the harness changes.
waiting…
waiting…
waiting…
Open the box.
Most agent demos show the finished box: impressive output, blurry machinery. These working examples open it.
The difference is the harness around the model.
Layer 1
The raw model.
A model turns text into text. On its own, it cannot see your orders, suppliers, or systems. It has no memory and cannot act.
Run it. Notice what it cannot know.
prompt:Walk through PO #12345. What's the status, what's blocking it, what's the next action?
Press Run to send the prompt to a raw model with no harness around it.
What just happened.
It answered from the prompt alone. It may admit what it cannot access or invent plausible details.
Every lookup, API call, and action comes from software around the model. The model proposes. The harness does.
Layer 2
The light harness.
A light harness adds tools and a loop: choose a tool, run it, return the result, repeat.
This is how many copilots and chatbots work. It is useful, and it is where many products stop.
prompt:Walk through PO #12345. What's the status, what's blocking it, what's the next action?
Press Run to send the prompt to a model with a single tool (`get_po`) wired up.
What just happened.
The model chose the tools. The harness ran them and returned the results. At this layer, the harness is a registry and a loop.
It can act, but each conversation still starts with whatever context fits in the prompt.
Layer 3
The real harness.
A real harness also carries a model of the business: purchase orders, suppliers, parts, work orders, and their relationships. We call those dimensions.
When a question crosses the operation, the harness finds the relevant context and coordinates multiple steps with shared state.
Run the same PO question. Watch it follow five connected nodes and build a next action.
prompt:Walk through PO #12345. What's the status, what's blocking it, what's the next action?
Press Run to send the prompt to a small simulated harness with a five-node dimension graph.
What just happened.
It moved from the PO to supplier history, the deadline, the part, and an alternate source, then built a recommendation.
The distinction is not more eloquent text. It is context construction, graph traversal, and coordinated action against a model of your business.
What most “agents” actually are.
Most agents are layer 2: a model with a tool loop. That is useful for bounded work.
For work that crosses the operation, ask what the harness knows, what it can access, and whether it keeps what it learns.
myai is layer 3. Its shared operational context persists across conversations, users, and apps. The next use case starts ahead.
Understanding precedes execution. Get the context right, and agents, apps, reports, and answers flow from it.