An AI agent is a language model wrapped in a loop that lets it take actions. It reads a request, decides on a step, calls a tool, reads the result, and repeats until the task is done or it needs you. The loop and the tools are what separate an agent from a chatbot, which only produces text.

What is the difference between an AI agent and a chatbot?

A chatbot responds. An agent acts. A chatbot's only output is text; an agent can call tools that change the world — sending an email, editing a file, running a command — and it keeps going for several steps without being prompted each time.

MCP stands for Model Context Protocol, an open standard for connecting AI agents to external tools and data sources. An MCP server exposes a set of capabilities, and any MCP-compatible agent can use them without custom integration code. It is the reason a new integration can work across many agents at once.

What is computer use?

Computer use is the capability that lets an agent operate a computer the way a person does: looking at the screen, moving the pointer, clicking, and typing. It is the fallback for software that offers no API, and it is slower and more error-prone than a direct integration, which is why well-designed agents try connectors first.

What is an approval gate?

An approval gate is a checkpoint where an agent stops and waits for a human before taking an irreversible or outbound action — sending a message, deleting a file, making a purchase. It is the single most important safety property of a personal agent, because it bounds the damage any mistake can do.

What is a surface, and what is handoff?

A surface is a place you can reach your assistant: messaging such as iMessage or WhatsApp, a mobile app, the web, or a desktop app. Handoff is continuing one task across two surfaces without re-explaining — starting it in a text and finishing it at your Mac. Handoff requires shared memory across surfaces rather than a separate conversation per channel, which is why most multi-channel assistants have channels but no handoff.

A Personal AI Agent Glossary: 61 Terms, Defined Plainly

Most glossaries in this space are padded. This one isn't. Every entry is something you'll actually run into while choosing or using a personal AI agent, defined in a sentence or two you could say out loud.

The basics

AI agent. A language model wrapped in a loop that lets it take actions. Read the request, pick a step, call a tool, read the result, repeat. The loop and the tools are the whole difference between an agent and a chatbot.

Agentic. An adjective for software that acts rather than just answers. Mostly a marketing word now, but the useful test behind it is real: can it do something you'd have to undo?

Chatbot. A model that produces text and nothing else. Useful, and not the same thing.

Personal AI agent. An agent working for one person on their own data and accounts, as opposed to a customer-facing bot or a team workflow system.

Assistant. In this category, usually a personal agent with a friendly face on it. There's no technical distinction.

Autonomy. How many steps an agent takes before it checks with you. Higher isn't better. The right level is "as far as it can go without doing something you'd want to reverse."

How agents work

Agent loop. Perceive, plan, act, observe, repeat. Every agent is some version of this, however it's dressed up.

Tool. A function the model can call — search the web, read a file, send an email, run a shell command. Tools are what make an agent useful and also what make it dangerous.

Tool call / function call. One invocation of a tool, with arguments the model chose. A good agent shows you these before running them.

Orchestration. Coordinating multiple steps, tools, or agents toward one outcome.

Planning / task decomposition. Breaking a big request into steps. Where a lot of agent failures start, because a bad plan executes perfectly.

Reflection. The agent checking its own output and trying again. Costs tokens, catches real mistakes.

Multi-agent system. Several specialised agents working together, usually with one coordinating. Powerful, and roughly twice as many ways to fail.

Agent handoff. One agent passing work to another, with the context it needs to continue. (Not to be confused with surface handoff below, which is about you moving between devices.)

Context

Context window. The maximum amount of text a model can hold at once, measured in tokens. Everything the agent knows in a given moment has to fit.

Token. A chunk of text, roughly three-quarters of a word. The unit both context windows and pricing are measured in.

Context engineering. Deciding what goes into the context window and what stays out. The highest-leverage skill in building agents, and mostly a subtraction problem.

Grounding. Giving the model real source material so its answers are anchored to something checkable rather than produced from training data.

RAG (retrieval-augmented generation). Fetching relevant documents and putting them in the context before the model answers. Good for knowledge. Not the same as memory.

Screen context. What the agent can see of your display right now. Cheap, immediate, and only useful if the agent asks before taking it.

Ambient context. The things the agent knows without being told — which app is in front, what you've selected, which tabs are open.

Memory

Memory. Anything the agent retains across conversations. Not the context window, which empties.

Short-term / working memory. What's in the current context window. Gone when the session ends.

Long-term memory. Persisted facts and preferences that survive sessions.

Memory channel. A named, editable store of memory that agents can subscribe to — for example a channel for your standing instructions and another for a specific project. The point of naming and separating them is that you can open one and fix it. See Memory and Channels.

Episodic memory. What happened. Semantic memory. What's true. Most products blur these; the distinction matters when the agent confidently tells you something that was true.

Memory hygiene. Pruning what the agent remembers. Unmanaged memory degrades into a pile of stale facts that quietly poison every answer.

Models

LLM (large language model). The model underneath. Claude, GPT, Gemini, Llama, and so on.

Provider. Where models come from — Anthropic, OpenAI, Google, a hosting service, or your own machine.

Model routing. Sending different jobs to different models: a fast cheap one to rewrite a sentence, a reasoning model for a hard plan. Saves real money and real time.

Local model. A model running on your own hardware. Private, free per token, and slower. See Models and Providers.

Reasoning model. A model that spends extra computation thinking before answering. Better at planning, worse at latency, more expensive.

Multimodal. Handles more than text — images, screenshots, audio.

Vision. The specific ability to interpret images and screenshots, which is what makes screen context possible.

Doing things in the world

Computer use. Operating a computer as a person would: looking at the screen, moving the pointer, clicking, typing. The fallback when there's no API, and slower and more fragile than one.

Browser use. The same idea confined to a web browser. Narrower, far more reliable.

MCP (Model Context Protocol). An open standard for connecting agents to external tools and data. Write an MCP server once and any compatible agent can use it. See MCP Tools.

Connector / integration. A pre-built link to a specific service. Safer and more limited than direct system access.

Skill. A reusable package of instructions that teaches an agent a recurring workflow, which can activate automatically based on the app you're in or what you asked. See Skills.

Gateway. A bridge that lets you reach your agent from messaging platforms — Slack, WhatsApp, iMessage, Telegram, email — instead of only from the app.

Scheduled task. An agent run that fires on a timer rather than on a request. A morning briefing is the canonical example.

Surfaces and continuity

Surface. A place you can reach your assistant: messaging (iMessage, WhatsApp, SMS), a mobile app, the web, or a desktop app.

Channel. One specific route within a surface — iMessage and WhatsApp are two channels of the messaging surface. Vendors often count channels and call them surfaces.

Handoff. Continuing one task across two surfaces without re-explaining. Requires shared state, not just multiple entry points. Both Claude Cowork and ChatGPT Work now do this within their own apps.

Chief of staff. A cloud-hosted personal agent you reach by message — always on, independent of whether your computer is open.

Co-work. A desktop agent working alongside you inside the applications you already have open, rather than in its own window.

Safety and control

Approval gate / human-in-the-loop. A checkpoint where the agent stops and waits for you before an irreversible or outbound action. The single most important safety property of a personal agent, because it caps the damage of any single mistake.

Permission scope. Exactly what the agent is allowed to reach. Ask every vendor for theirs in writing.

Sandboxing. Running the agent's work in an isolated environment so mistakes can't touch your real system.

Prompt injection. An attack where hostile text hidden in a web page, email, or document issues instructions the agent follows. The defining security problem of this category, and not solved.

Data exfiltration. An agent being tricked into sending your data somewhere it shouldn't. Usually the payload of a successful prompt injection.

Damage radius. How much harm one bad agent decision can cause. Designing for a small radius beats hoping for good decisions.

Hallucination. The model producing something confident and false. Agents make this worse, because a hallucination can become a tool call.

Evaluation

Eval. A repeatable test of agent behaviour. Without one you're guessing.

Success rate. How often the agent completes the task. The headline metric, and insufficient on its own.

Benchmark. A standard task suite run across products. In this category, Assistant Benchmark scores personal assistants on the same 15 tasks.

Trace. The full record of an agent run: every thought, tool call, and result. The only practical way to answer "why did it do that?"