Ben ShapiroRobot in a Room

Robot in a Room

Fourteen steps, one at a time, until a chatbot has become an agent.

Most people meet AI as a chat window. Then the words pile up: context window, tools, MCP, skills, agents. This is for anyone who uses these tools at work and wants to know what each word adds. One robot in one sealed room gets one new thing at each step, and every step ends with where the picture stops being true.

Ben Shapiro · 10 November 2026 · About ten minutes

Scroll to begin, or press →

The 3D room could not start in this browser, so here is the full text on its own. Every step is below.

Step 1 of 14

The room

Real nameLanguage model

A language model is a robot sealed in a room. A written message comes in through one slot. The robot reads it, writes a reply and pushes it out the other slot. Text in, text out. That is all it does.

Where the picture bends

The robot writes its reply one piece of a word at a time, each a prediction from patterns it learned in training. It does not look anything up.

First shipped

November 2022: ChatGPT put this room in front of the public.

Step 2 of 14

The desk

Real nameContext window

The desk is the only place the robot can look. Your message, its earlier replies, any document you hand over: all of it has to fit on the desk. The desk has a fixed size. When it fills, the oldest pages fall off.

Where the picture bends

Desk space is counted in tokens (pieces of words), not pages. A crowded desk is also harder to work at: details get missed as it fills.

First shipped

There from the start, and growing: from about 4,000 tokens in 2022 to 1 million tokens, previewed in February 2024 (Google’s Gemini 1.5 Pro).

Step 3 of 14

A notice on the wall

Real nameSystem prompt

Before any message arrives, whoever runs the building pins a notice to the wall: who the robot is, how to behave, what to avoid. The robot reads it before every reply. You rarely see the notice, but it shapes every answer.

Where the picture bends

There is no wall. The notice is the first page on the desk, so it uses desk space too. It guides the robot. It does not bind it.

First shipped

Early 2023, when OpenAI’s chat models reached developers with a separate “system” message.

Step 4 of 14

Scratch paper

Real nameReasoning

Hard questions go better with working-out. Give the robot scratch paper and it drafts, checks and corrects before it answers. The scratch work stays in the room. Only the finished reply goes out through the out slot.

Where the picture bends

Scratch paper sits on the desk and uses space like everything else. It also costs time and money. What you are shown of it is often a summary.

First shipped

2024, with OpenAI’s o1 models, the first widely available ones built to reason before they answer.

Step 5 of 14

A wider slot

Real nameMultimodal input and output

Widen the slots and more than text fits through: photos, charts, scanned pages, audio. That is what “multimodal” means: more than one kind of content. Some models can send images and sound back out as well.

Where the picture bends

Everything is still turned into tokens before it reaches the desk, and pictures are bulky. One image can take the space of several pages of text.

First shipped

September 2023: image input began rolling out in ChatGPT.

Step 6 of 14

A phone

Real nameVoice

Put a phone in the room and you can simply talk. Early versions used a typist: your speech was typed up for the robot and its reply was read aloud. Newer models hear and speak directly, so they are quicker and catch tone.

Where the picture bends

It is the same robot at the same desk. Voice changes how the message arrives, not what the robot knows.

First shipped

September 2023: voice conversations in ChatGPT, the typist version. Models that hear and speak directly followed in 2024.

Step 7 of 14

A filing cabinet

Real nameMemory

The desk is cleared after every conversation. A filing cabinet lets notes survive: your preferences, project facts, last week’s decisions. When a new conversation starts, the relevant notes are pulled out and put on the desk.

Where the picture bends

The robot does not remember as you do. It knows only what was written down and fetched this time. Notes can be stale, or missed. Most products let you read and delete them.

First shipped

2024, when memory arrived in ChatGPT. Other assistants followed.

Step 8 of 14

A bookshelf

Real nameSkills

A shelf of binders, each a written procedure for one job: build the quarterly report, apply the brand. The robot sees only the labels on the spines. When a job matches a label, it pulls that binder and reads it at the desk.

Where the picture bends

“Skills” started as Anthropic’s term. Since December 2025 it is an open format that other tools read too. A binder holds instructions and files, not new ability. The robot is no smarter, only better briefed.

First shipped

October 2025: Anthropic introduced Agent Skills for Claude. It opened the format as a standard that December.

Step 9 of 14

A tool hatch

Real nameTool use

The robot cannot leave the room, but it can ask the building to act. It writes a request slip, such as “search the web for this”, and passes it up through the hatch. The building runs the tool and sends the result back down. The robot never holds the tool.

Where the picture bends

There is no hatch. A slip is a specially formatted part of the reply, which the surrounding software spots and acts on. That software decides which tools exist.

First shipped

June 2023: OpenAI added “function calling” to its API. Every major model now has it.

Step 10 of 14

Cables to your apps

Real nameConnectors, built on MCP

Wiring in each tool by hand is slow, so the industry settled on a standard plug: the Model Context Protocol, or MCP. Run one cable to an app, such as your email or CRM, and it supplies several tools at once: search, read, draft, update.

Where the picture bends

“Connector” is Claude’s product name. Other products say app or integration. The standard underneath is open. A cable carries your sign-in, so the robot reaches only what that account can.

First shipped

November 2024: Anthropic published MCP as an open standard. Other major AI vendors adopted it during 2025.

Step 11 of 14

A workshop

Real nameSandbox for running code

Some jobs need more than words: crunch a spreadsheet, draw a chart, convert a file. The workshop is a sealed side room, a “sandbox”, where code is run, so answers are calculated and not guessed. With “computer use”, the robot can also operate a screen, mouse and keyboard.

Where the picture bends

The robot never walks in. It passes code through the window; the building runs it and hands back the output. Computer use is the same routine: a screenshot comes back, the next click goes in.

First shipped

2023, as Code Interpreter in ChatGPT. Computer use followed in late 2024, as a beta from Anthropic.

Step 12 of 14

Working in rounds

Real nameThe agent loop

Nothing new is added. The robot is allowed to keep going. It thinks, sends a slip, reads the result, decides the next step, and goes around again, as many times as the job needs. A robot working this way is called an agent.

Where the picture bends

Every round adds pages to the same fixed desk, so long jobs need tidying along the way. A loop also needs a stop rule: job done, budget spent, or a person stepping in.

First shipped

The pattern comes from research in the early 2020s. It became a mainstream product feature around 2025, led by coding agents.

Step 13 of 14

A checkpoint

Real nameHooks and permissions

A robot that can act needs limits that do not rely on its judgment. So the building installs a checkpoint. Permissions are the rule list: allowed, ask a person first, never. Hooks are fixed steps the building runs at set moments. They run every time. The robot has no say.

Where the picture bends

“Hooks” is the term in Anthropic’s Claude Code. Elsewhere, similar controls are called guardrails or approvals. The notice on the wall asks. The checkpoint enforces.

First shipped

Approval prompts shipped with the first coding agents. Hooks reached Claude Code during 2025.

Step 14 of 14

More rooms

Real nameSubagents

Big jobs get split. The lead robot writes a brief and passes it to another room, where a second robot starts at its own empty desk. Several can work at once. Each sends back only a short report, which keeps the lead’s desk clear.

Where the picture bends

A helper knows only what its brief says. It never saw the lead’s desk. Vague briefs get vague work, and every extra room costs money to run.

First shipped

2025, when Claude Code added custom subagents. Research systems had tried teams of agents earlier.

Terms that get mixed up

Five of these words are used loosely, sometimes for each other. Each one adds something different to the room.

Tool
One action the building can carry out for the robot: search, send, look up, calculate. The robot asks for it on a slip.
Adds reachAlso called a function or an action.
Connector
A cable to one app that supplies a bundle of tools, using the open MCP standard.
Adds a source of toolsA product name. Elsewhere: integration, app, or MCP server.
Skill
A binder of written instructions for doing one kind of job well. Opened only when the job calls for it.
Adds know-howStarted at Anthropic, now an open format. A skill may tell the robot which tools to use, but it is not a tool.
Memory
Notes kept in the cabinet between conversations and put back on the desk when relevant.
Adds continuityNot the same as the desk, which is cleared every time.
Hook
A fixed step the building runs at a set moment, whatever the robot thinks about it.
Adds controlClaude Code’s term. Elsewhere: guardrail, policy, approval step.

In short: tools let the robot act, connectors bring tools in from your apps, skills tell it how to do a job, memory carries notes between conversations, and hooks are steps the building runs whatever the robot decides.

The building has a real name too. It is the software around the model, which runs the tools, the cables and the checkpoint. Engineers call it the harness. Most people call it the app.

Every step, and its real name

In the roomReal nameWhat it adds
01The roomLanguage modelReads text and writes text. Nothing else.
02The deskContext windowEverything the model can see at one time.
03A notice on the wallSystem promptStanding instructions from whoever runs the system.
04Scratch paperReasoningWorking-out before the answer.
05A wider slotMultimodal input and outputImages, audio and files, not only text.
06A phoneVoiceSpoken conversation in both directions.
07A filing cabinetMemoryNotes that survive between conversations.
08A bookshelfSkillsWritten procedures, opened only when needed.
09A tool hatchTool useA way to ask outside software to act.
10Cables to your appsConnectors, built on MCPA standard plug that brings an app’s tools with it.
11A workshopSandbox for running codeA sealed place to run code and operate a computer.
12Working in roundsThe agent loopPermission to keep going, round after round.
13A checkpointHooks and permissionsRules the software enforces every time.
14More roomsSubagentsHelpers with their own robots and their own desks.

Next: The Factory Floor. Pull back, and this room is one workstation on a production line.