AI IS JUST
AN OFFICE.
No magic. No jargon. Just employees, desks, filing cabinets, phones—and one surprisingly helpful coffee machine.
↓ Seriously, none.
ME!
IDEAS
MAGIC!
THE OFFICE TOUR
Let's meet
the whole team.
Scroll through each room. Poke things. Ring phones. Make a mess. By the time you reach the lift, modern AI will feel oddly familiar.
↓The Claude app is the front desk.
When you type into Claude, the app receives your message and hands it to the model you selected. For this tour, that employee is Claude Opus 4.8. You never need to call Opus directly.
ReceptionA chat app is a user interface and orchestration layer. It packages your input, applies the selected settings and capabilities, sends a request to a specific model, then displays that model’s output.
Opus is the brilliant employee.
The Claude app at reception passes your message to Opus 4.8, then carries its answer back. You talk to the front desk; this named model does the thinking behind it.
Meet Opus 4.8Claude Opus 4.8 is a specific large language model in Anthropic’s Claude family. An LLM is a neural network trained on large datasets to process and generate token sequences; Opus is the model doing that work in this example.Opus can only use what fits on its desk.
It's the same employee from the previous room. A bigger desk lets Opus hold more instructions, examples, and conversation at once.
Context windowThe context window is the maximum number of tokens Opus can process at once—including instructions, chat history, retrieved documents, tool results, and generated output.Memory is the archive beside Opus.
It is still the same Opus employee. An outside memory system stores useful details, retrieves them when needed, and puts them back on Opus's desk. The receptionist does none of this.
MemoryAI memory is usually application data stored outside the model. The application or agent harness retrieves a useful detail and inserts it into a future context; the model itself does not continuously remember it or change its weights.Opus chooses. The harness operates.
The same Opus employee decides a calculator would help. The blue office runner—the application or harness—operates it and brings the result back. The Claude app receptionist is not the tool user.
ToolsA tool is a callable function or external operation described to the model with a name, purpose, and input schema. Opus can produce a structured tool request, but the host application or agent harness validates and executes it, then inserts the result into Opus’s context.
Is it a harness? That's a useful mental model. Anthropic calls Claude Code an agentic coding tool: the selected model—Opus 4.8 in this tour—is the employee; Claude Code is the equipped coding workplace wrapped around it.
Opus asks. The researchers fetch.
Opus decides what it needs, then a browser runner searches the wider world or a RAG librarian fetches from a trusted collection. The sources return to the same Opus employee's desk. Perplexity and NotebookLM are familiar examples.
Browser + RAGBrowsing and retrieval happen outside the model. A browser tool searches external sources; a RAG system retrieves relevant documents—often using embeddings and similarity search. The application then adds the returned material to Opus’s context before generation.
FOLDER ✓
The harness makes the call for Opus.
Opus requests a clearly defined action. The blue runner calls the outside service, then puts its structured answer on Opus's desk. Stripe's payments API and Twilio's communications API are famous examples.
APIsAn API defines endpoints, request formats, authentication, and response formats so software systems can communicate predictably. In an AI workflow, the host application or harness typically sends the network request—not the model itself—and returns the response as tool-result context.
MCP gives the harness standard doorways.
Different products agree on the same kind of entrance. Opus selects an available capability; the blue harness runner goes through the MCP doorway and brings the result back. GitHub, Slack, Notion, Figma, and Google Drive are familiar destinations.
MCPModel Context Protocol is an open client-server standard. A host application or agent harness acts as the MCP client, discovers capabilities from MCP servers, exposes allowed tools or resources to the model, executes calls, and returns results to the model’s context.
One employee is useful.
A team is unstoppable.
An agent harness is the manager and equipped office around a model such as Opus. The model reasons inside each step; the harness holds the goal, loops, runs tools, checks results, and decides when to stop. A team can use multiple model instances with different roles. Claude Code and OpenAI Codex are familiar examples.
Agents + harnessesAn agent is a software loop in which a model selects actions, uses tools, observes results, and continues toward a goal. The harness manages that loop, including context, permissions, state, retries, and stopping conditions.
☑ ☑ □
Fine-tuning is specialist training.
Start with a capable general employee. Train them on many examples of one kind of work, and they become more consistent at that specialty.
Fine-tuningFine-tuning continues training a pre-trained model on selected examples so its weights adapt toward a task, domain, response style, or behavior.+ feedback = specialist
Companies build the employees.
OpenAI, Anthropic, Google, Meta, Mistral, and DeepSeek each train their own kinds of AI employees. The employee is the model. The company is the builder.
Different buildings.Different employees.
Same city.
THE NO-NONSENSE GLOSSARY
AI speak,
translated.
Click a piece of jargon for the office translation. Hover, focus, or tap its Tech note for the real definition.