How Does the Agentic Stack Fit Together? Model, Agent, Skills, Tools, Orchestration
The agentic stack has five layers, and each maps to something you already run in your business: the model is the rented brain, the agent is the worker built around it, skills are the written playbooks, tools are the hands that touch your systems, and orchestration is the manager coordinating multiple workers. Once you hold that picture, every vendor pitch sorts itself into "which layer are they actually selling?"
The glossary defines each of these terms in a line or two. This is the article that snaps them together — because in the wild, they never arrive one at a time. They arrive in a demo, stacked and blurred, and the person who can't tell the layers apart is the person who overpays. Here's the whole machine, top to bottom.
Layer 1: The model — the rented brain
At the bottom sits the large language model: Claude, GPT, Gemini. It's the raw intelligence — the thing that reads, reasons, and writes. Three facts about it matter to a buyer:
- Almost nobody builds one. Training a frontier model costs on the order of what a mid-sized company is worth. You rent intelligence the way you rent cloud compute — per token, from Anthropic, OpenAI, or Google.
- It's stateless. The model has no memory of your business between requests. Everything it "knows" in a given moment was placed in its context window for that moment — the discipline covered in context engineering.
- It does nothing on its own. A model is a brain in a jar. It can't check your email or update your CRM. Every capability beyond text comes from the layers above.
So when a vendor implies proprietary intelligence, the grounded question is: which model is underneath? The answer is almost always one of the big three, and that's fine — the vendor's real value lives upstairs.
Layer 2: The agent — the worker built around the brain
An agent is what you get when you wrap a model in a loop: take a goal, plan the steps, act through tools, look at the result, adjust, repeat until done. The model supplies judgment at each turn of the loop; the agent structure supplies persistence. That loop is the entire difference between AI that answers and AI that works — unpacked fully in what is agentic AI.
The business analogy: the model is raw talent, the agent is that talent given a desk, a login, and an assignment. Same brain, employable form.
Layer 3: Skills — the playbooks your business writes
Skills are documented procedures an agent loads when a job calls for them: how your company qualifies a lead, formats a proposal, reconciles an invoice. If the model is talent and the agent is the worker, skills are your SOPs — the difference between a temp who improvises and a trained employee who does it your way, every time.
This layer is easy to overlook and strategically the most interesting, because it's the one you own. Models are rented, agent frameworks are increasingly commodity, but the encoded knowledge of how your business runs — that's yours, it compounds, and it transfers when you swap models or platforms. A founder who spends a week writing sharp playbooks gets more from a mediocre agent than a founder with no playbooks gets from a frontier one. (Codified, reusable playbooks are the entire premise of the Optimus Frameworks library.)
Layer 4: Tools — the hands on your systems
Tools are the specific capabilities an agent is allowed to use: read email, query the database, run code, browse the web, write to the CRM. No tools, no action — an agent without tools is just a chatbot with ambition.
Two buyer-relevant points live here:
- Tools define the blast radius. The tool list is the permission list. An agent that can only read your calendar can't do much damage; an agent that can send external email needs an approval step. When you hear "guardrails," this layer is mostly where they're bolted on.
- MCP lives here. The Model Context Protocol is a standard socket for plugging tools and data sources into agents, so every integration isn't a custom build. "MCP support" in a pitch means "connects to the agent ecosystem the standard way" — useful, real, and not magic.
Layer 5: Orchestration — the manager of many workers
Orchestration is what appears when one agent isn't enough: a coordinating layer that breaks a big job into pieces, assigns them to specialized agents, sequences the handoffs, and assembles the result. One agent researches, another drafts, a third reviews — orchestration is the manager keeping them from stepping on each other.
It's the most impressive-sounding word in the stack, which is why it shows up in pitches for products that are, on inspection, a single chatbot. Honest rule of thumb: if there's only one worker, there's nothing to orchestrate. Most $5–50M businesses don't need this layer on day one — they need one agent doing one job well, then a second. The manager gets hired when there's a team.
The whole stack on one table
| Layer | Plain English | Business analogy | Who owns it |
|---|---|---|---|
| Model | The intelligence | Rented brain / raw talent | Anthropic, OpenAI, Google |
| Agent | The loop that gets work done | The worker at the desk | Platform or your builder |
| Skills | Documented procedures | Your SOPs and playbooks | You |
| Tools | Permitted capabilities | Hands, logins, permissions | You grant; vendors connect |
| Orchestration | Coordination of many agents | The manager | Platform, when you're ready |
How do you use this picture as a buyer?
Run every pitch through one question: which layers did this vendor actually build, and which are they reselling? "Our proprietary AI" usually decodes to "a rented model plus our agent loop and integrations" — which can absolutely be worth paying for, but should be priced as scaffolding, not as a private frontier model. And the layer where your leverage lives — skills, the encoded way your business runs — is the one no vendor can sell you and no vendor can take with them when you leave. Fluency across all five layers takes an evening. The vocabulary for the rest of the pitch meeting takes twenty minutes — that's what the glossary is for.
FAQ
What's the difference between a skill and a tool?
A tool is a capability — something the agent can do, like send email or query a database. A skill is a procedure — how your business wants a job done, step by step, written down so the agent follows it the same way every time. Tools are the hands; skills are the playbook the hands follow.
Do I need orchestration to start with agents?
No. Orchestration matters when multiple agents hand work to each other. Most businesses should start with one agent, a handful of tools, and one well-documented workflow — and add the manager layer only when there's actually a team of agents to manage.
Where does MCP fit in the stack?
MCP (Model Context Protocol) lives at the tool layer. It's a standard way to plug tools and data sources into an agent — a universal socket, so every integration doesn't have to be custom-built. When a vendor says "MCP support," they mean their product can connect to agents through that standard.
Which layer matters most when evaluating an AI product?
The layers the vendor actually built. The model is usually rented from Anthropic, OpenAI, or Google — so the vendor's real value lives in the agent loop, the tool connections, and the skills. Ask which model is underneath, what tools connect out of the box, and how your procedures get encoded.