How to Make Sense of Harness, Agent, Tool, Model, MCP Server, and Skill

If you follow modern AI development, you’ve likely seen terms like Model, Agent, Harness, and MCP Server thrown around interchangeably. While they work together closely in modern AI systems, mixing them up makes it hard to understand how AI applications actually function.

Think of building an AI application like setting up an automated workshop: you need a brain to process information, hands to take action, plugs to connect to remote systems, a playbook to follow, a worker with a clear goal, and a manager to coordinate the entire environment.

Here is how each piece fits into the puzzle.

Harness  — the stage and manager
  │
  ├─ Agent  — a goal-directed worker
  │    loop: Observe → Decide → Act  (until a stop condition)
  │    │
  │    ├─ Model     the brain     (predicts tokens / tool calls)
  │    ├─ Context   working memory for this run
  │    ├─ Tools     the hands     (native, or via an MCP server)
  │    └─ Skill     the playbook  (how to do a multi-step job)
  │
  └─ MCP Server  — the universal plug
       advertises remote tools, resources, and prompts

The rest of this note is one section per term: what it is, what it is not, and a concrete example where that helps.

1. Model: The Brain

What it is: An AI engine that predicts the next set of tokens—whether that is plain text or a structured request to call a tool.

What it isn't: It has no intrinsic goals, no execution loop, no filesystem access, and no memory beyond the specific text window (context) provided to it in the moment.

Example: Claude Opus 4.8

A model is extremely capable inside that window, and extremely limited outside it. If something happened, it happened because another piece of the system asked the model, fed it context, or ran what it proposed.

2. Tool: The Hands

What it is: A single, named function with a strict format for inputs and outputs. It gives an AI system the ability to perform specific actions.

Types:

• Native Tool: Built directly into the local application (e.g., a file-listing tool to inspect local folder contents).
• MCP-Backed Tool: Connected from an external service (e.g., creating a Pull Request on GitHub).

How it works: The model proposes using a tool, and the execution system runs it.

That last sentence is the whole split. The model does not “have tools” in the way a person has hands. It emits a request in a shape the harness understands. Something else is allowed — or not allowed — to actually execute it.

3. MCP Server: The Universal Plug

What it is: Model Context Protocol (MCP) is a standardized way of interfacing with capabilities running outside the local system. An MCP Server advertises available tools, data resources, and prompts to the local environment.

Example: An MCP server for GitHub that allows a local application to securely inspect and manage remote repositories.

MCP is not the tool itself, and it is not the model. It is the socket. Once a server is connected, the harness can offer those remote tools to an agent the same way it offers native ones.

4. Skill: The Playbook

What it is: A structured set of instructions, workflows, or procedural guidance that teaches an AI system how to perform a complex, multi-step job effectively.

How to think of it: If a tool is a hammer, a skill is the step-by-step woodworking guide explaining how to build a cabinet.

Skills do not replace tools, and they do not replace the model. They bias the worker toward a known good procedure: which tools to reach for, in what order, with what checks, and when to stop.

5. Agent: The Goal-Directed Worker

What it is: An active system combining a Model, Context (memory/information), Tools, and a Stop Condition. It operates in a continuous loop: Observe → Decide → Act until the assigned goal is achieved or an error occurs.

Subagents: Helper agents spawned with their own separate contexts to handle specific sub-tasks independently.

Example: An agent assigned to locate a file in a large codebase. It reviews project files, decides what search tools to use, checks the results, and stops once the file is found.

An agent is not a synonym for a model. A model without a loop, tools, and a stop condition is just answering one prompt. An agent is the worker who keeps going until the job is done — or until the harness says the job is over.

6. Harness: The Stage and Manager

What it is: The top-level runtime environment that brings everything to life. It instantiates agents, builds their context, calls the model, executes tool calls, enforces permissions, compacts conversation history, spawns subagents, and decides when the task is complete.

Example: Claude Code

If the agent is the worker, the harness is the shop: benches, power, safety rules, and the manager who decides who is on shift. Most of the behavior people attribute to “the model” in a product is actually harness policy — what context gets loaded, which tools are visible, what a skill is allowed to instruct, and when the loop is allowed to continue.

Putting it together

A useful order to remember:

1. The model predicts.
2. A tool is a single action it can request.
3. An MCP server is how remote actions and data get plugged in.
4. A skill is how to do a longer job well.
5. An agent is the worker looping until a goal is met.
6. The harness is the environment that makes that loop real, safe, and stoppable.

Once those six stay distinct, the rest of the stack gets easier to read: you can tell whether a limitation is in the brain, the hands, the plug, the playbook, the worker, or the shop that runs them.