I often perceive group usage the words agent and model interchangeably, referring to Claude arsenic either one. So I thought it would beryllium useful to constitute a speedy reference connected the terminology we usage here, to thief america person much precise conversations.
Let’s commencement pinch a schematic that shows wherever we are headed.
The Agent System
Harness Inference service ModelAn supplier strategy is made up of respective layers. At its halfway is simply a model. Things for illustration Sonnet, Opus, aliases Gemini. These are trained connected immense amounts of matter and data, and successful the end, they are fundamentally large collections of floating constituent numbers wired together successful a peculiar way.
Frontier models are acold excessively computationally costly for astir of america to tally locally astatine afloat scale. They request measurement much RAM than astir of america person connected our section machines. So we request location other to tally them. That location is an inference service. Services for illustration AWS Bedrock aliases Anthropic’s API. The conclusion work takes your API calls, feeds them into the model, and besides tracks pricing arsenic you go.
The work runs the exemplary successful an conclusion engine, but it is still beautiful basic. Text in, matter out. Think of really ChatGPT worked erstwhile it first launched. That relationship layer, the point that gives you a bully measurement to talk to the API, is called a harness. In its simplest form, it is conscionable a lightweight wrapper. Other harnesses you mightiness cognize are Claude Desktop aliases Claude CLI.
And this is wherever things get interesting. Features for illustration MCP and Skills? They are chiefly portion of the harness layer. The exemplary doesn’t inherently cognize astir an MCP server aliases a Skill; the harness decides what discourse and devices to expose to it.
So if you put each that together, an agent system is simply a harness, a group of devices and logic for processing inputs, that calls an conclusion service, which runs a model. That is it. That is the full stack.
Real world examples
Here is really the stack breaks down for immoderate communal devices you mightiness beryllium using.
| Claude Desktop | Claude Desktop (UI + MCP + section logic) | Anthropic’s conclusion service | Sonnet / Opus / Haiku |
| Claude CLI | Claude CLI (tool parsing + record I/O) | Anthropic’s conclusion service | Sonnet / Opus / Haiku |
| Cursor | Cursor editor (context assembly + instrumentality routing) | Cursor’s conclusion furniture (various providers) | Sonnet / GPT / Gemini / etc. |
| ChatGPT | ChatGPT UI (history + orchestration) | OpenAI’s conclusion service | GPT models |
| Custom supplier you build pinch LangChain | Your LangChain code (prompt templates + instrumentality definitions) | Your chosen supplier (Bedrock, OpenAI, etc.) | Your chosen model |
Notice the pattern. The harness is wherever your logic lives. The conclusion work is the hosted furniture that runs the model. The exemplary is the mathematical point that produces text. The aforesaid model, opportunity Sonnet, tin beryllium utilized crossed aggregate supplier systems pinch wholly different harnesses, and it will behave otherwise because the harness is shaping the inputs and interpreting the outputs.
Let’s ideate we are building a house. We person a building unit connected site. They return a blueprint, bid materials, grip equipment, and determine sequencing. They are the only ones who tin really touch the ground. Pour concrete, hammer nails, that benignant of thing. But if thing comes up and they request brainpower, they telephone an architect. They can’t talk to the designer directly, though. They person to spell done the patient that employs them. The patient handles scheduling and billing. And the designer is very particular. You springiness it a brief, it gives you backmost insubstantial plans. Nothing more.
The unit is the harness. They’re the portion that tin really touch the extracurricular world and move the architect’s plans into actions. The patient is the conclusion service. The gateway that handles logistics and cost. And the designer is the model. Pure, constrained, and superb astatine its constrictive job.
So erstwhile you usage thing for illustration Claude CLI, the CLI is the harness. It uses Anthropic’s conclusion service, which runs their models, Sonnet and Opus. One absorbing implication: arsenic models get smarter, they mightiness make immoderate of today’s harness logic, for illustration Skills aliases MCP, little useful. The measurement we build harnesses now mightiness not property well.
The takeaway
Let’s beryllium definitive pinch our terms.
- Model - the mathematical usability that transforms input tokens into output tokens.
- Inference service - the hosted work that runs the exemplary and tracks usage.
- Harness - the logic that shapes inputs, interprets outputs, and touches the extracurricular world.
- Agent system - each 3 moving together.
When we opportunity “my exemplary is doing this aliases that,” we are usually talking astir what the harness is orchestrating. The models themselves are conscionable these inscrutable mathematical objects that we get to telephone retired to.
And that favoritism matters. Because if thing goes wrong, aliases if we want to make things better, we request to cognize wherever to look. Is the exemplary giving bad answers? Maybe it needs amended discourse from the harness. Is it excessively slow aliases excessively expensive? That is astir apt the conclusion work aliases the compute underneath. Is it not utilizing devices correctly? The harness is astir apt formatting them incorrect aliases not parsing the responses properly.
When you tin sanction the layer, you tin hole the layer. That is the full constituent of being precise. It is not astir being pedantic. It is astir being capable to amended things faster and much effectively.
| Bad reasoning / knowledge | Model aliases discourse supplied by harness |
| Missing context | Harness |
| Tool isn’t available | Harness / instrumentality integration |
| Tool telephone is malformed | Harness aliases model |
| Tool executes incorrectly | Tool / harness |
| Slow inference | Inference infrastructure |
| High cost | Model prime / conclusion service |
| Same exemplary behaves differently | Harness / discourse / tooling |
English (US) · 
Indonesian (ID) · 
