Model, app, harness, tool: the AI vocabulary that separates a PM from an engineer

Why four words are worth the trouble

In nearly every training session I run, the same thing happens in the first hour. Someone says "the AI keeps doing X," someone else pushes back, and after a few minutes it's clear they were picturing completely different things. One meant the raw model. One meant the chat product. One meant an autonomous agent wiring itself into their systems. They weren't disagreeing about the facts. They didn't share the words.

Engineers already have these words; they just rarely say them out loud, because to them the distinctions are obvious. For a product manager, leader, or designer, learning the four layers is the single fastest way to stop talking past the technical team. It turns "can AI do this?" into a precise question about a specific layer.

The four layers, using the Claude stack

I teach this with one ladder, because a concrete example sticks better than a definition. Here is the same product, climbing from raw capability to something that acts on the world.

• Model: Claude. A model is a text predictor: given everything so far, it produces the next most-likely chunk. It has no memory between conversations, no access to your files, and a training cutoff. Powerful, but inert on its own.
• App: Claude.ai. An app wraps the model for a job to be done: a chat window, a place to upload a file, some memory, a system prompt that sets the behaviour. Same model, now usable by a human for a purpose. ChatGPT is the app; GPT is the model underneath.
• Harness: Claude Code. A harness gives the model a goal instead of a single prompt, then lets it loop: plan the work, call tools, read the result, decide the next step, and check whether it actually finished. This is the leap from "answers a question" to "completes a task."
• Tools: MCP servers. Tools are what the harness is allowed to reach: your codebase, a database, a calendar, an internal API. The Model Context Protocol is just a standard way to expose those tools so any harness can use them safely.

Read the ladder back the other way and the dependencies are obvious: a tool is useless without a harness to call it, a harness is just a loop without a model to drive it, and a model is a clever parlour trick until an app puts it in front of a real person.

Workflows and agents are not the same thing

The harness layer is where most of 2026's confusion lives, so it's worth one more distinction. Anthropic's own engineering team draws a clean line between workflows and agents. In a workflow, you (the human) wire the steps together in a fixed order, and the model fills in each box. In an agent, the model decides the steps itself, in a loop, until it judges the job done.

That difference is the whole conversation hiding behind the word "agentic." Most products people call agents today are really orchestrated workflows with a model in a few of the boxes, which is often the right, safer choice. Knowing which one you're building changes how you test it, how much you trust it, and where you put a human checkpoint.

How to actually use the vocabulary

You don't need to memorise this to get value from it. You need to use it to locate a disagreement. The next time a feature debate goes in circles, ask which layer everyone is talking about:

• "Is the limitation in the model, or in how our app is prompting it?", usually it's the second, and that's a thing you can fix this week.
• "Are we building a workflow or an agent here?", a far better question than "is this AI or not."
• "What tools does this harness actually need, and which ones should it never be allowed to touch?", this is a security conversation, and it's where most of the real risk lives.

Teams that share this vocabulary scope features faster and argue less, because they're finally pointing at the same thing. It's the first hour of every course we run for exactly that reason, and the cheapest upgrade a non-technical team can make to how it talks about AI. If you want your whole team fluent in it, that's what hands-on AI training is for; if you're not sure where your people stand today, the free AI proficiency assessment gives you a baseline in a few minutes.