Glossary

AI agent terms, defined

Plain definitions of the words used to describe AI agents and assistants — what each one actually means, and why it matters when you are choosing between products.

AI agent

An AI agent is software that pursues a goal by taking actions — browsing sites, filling forms, calling tools, waiting and retrying — rather than only producing text.

The distinguishing feature is not intelligence but scope of responsibility. A chatbot's loop is one turn long: you ask, it answers, and acting on the answer is your job. An agent's loop stays open until the goal is met or it hits something only you can resolve.

In practice an agent is a model plus tools, memory, a scheduler, and guardrails. Remove the tools and you have a chatbot with extra steps. See what is an AI agent for the longer treatment.

Chatbot

A chatbot converts a question into an answer and stops there; anything that needs doing afterwards is left to the person who asked.

Chatbots are not a lesser technology — for writing, analysis, and thinking out loud they are the right tool, and a good one outperforms any narrow agent at those jobs.

The limitation is structural rather than technical. A chatbot ends with you knowing something. Whether that knowledge becomes a booked table or a sent email is outside its remit.

Agentic browsing

Agentic browsing is when an AI system operates a real web browser — clicking, typing, and navigating pages the way a person would — in order to complete a task on a site that offers no API.

This matters because most of the web was never built for machines. Restaurants, ticket sellers, and local services have websites, not developer platforms. An agent that can only call APIs is limited to the small slice of the internet that publishes one.

It is also where agents fail most visibly: logins, payment walls, and CAPTCHAs stop a browsing agent cold, and a site that changes its layout can break a task that worked yesterday.

MCP (Model Context Protocol)

MCP is an open standard that lets an AI application connect to external tools and data sources through a common interface, so capabilities can be added without custom integration work for each one.

Before a shared protocol, every integration between an assistant and a service was bespoke. MCP makes the connection generic: a service exposes an MCP server, and any MCP-aware assistant can use it.

For someone choosing an assistant, MCP support is a proxy for whether the tool set is fixed or extensible — whether you are limited to a list someone else maintains.

Tool call

A tool call is the moment an AI model stops generating text and instead invokes an external function — a search, a calendar write, a browser action — then continues using the result.

Tool calls are the mechanism underneath every claim that an assistant 'did' something. The model decides which tool fits, supplies the arguments, and reads back what happened.

They are also where things go wrong quietly. A tool can fail, return nothing, or return something the model misreads, and a poorly built agent will summarise that as success.

Persistent memory

Persistent memory is durable storage of facts and preferences that an assistant carries between conversations, so context does not reset each time you start talking.

Without it, you re-explain yourself indefinitely and the assistant cannot improve at knowing you. With it, a request like 'book the usual' becomes meaningful.

The question worth asking of any product is whether the memory is inspectable. Memory you can open, correct, and delete is a meaningfully different proposition from personalisation you cannot see.

Context window

A context window is the amount of text a language model can consider at once, measured in tokens; anything outside it must be retrieved or summarised to be used.

Context windows are often confused with memory. The window is short-term working space for a single exchange. Persistent memory is what survives after it closes.

Monitor (standing instruction)

A monitor is a standing instruction to re-check something on a schedule and report only when a stated condition is met — a price falling below a threshold, an item coming back in stock, a slot opening.

A monitor has three parts: the thing to check, the condition that makes it interesting, and how long to keep going. Vague versions produce noise; 'under $500' is actionable where 'cheap' is not.

The quality signal is silence. A monitor that reports every check is a notification generator. See price and restock monitoring.

Proactive messaging

Proactive messaging is when an assistant starts the conversation — surfacing a result, a change, or a deadline without being asked first.

This is what makes background work useful. An agent that runs for hours needs somewhere to reach you, and a web app you have closed is not it.

It is also the feature most easily abused. An assistant that messages you to manufacture engagement is worse than one that stays quiet.

Confirmation gate

A confirmation gate is a required human approval before an agent takes an irreversible action such as spending money, sending a message, or cancelling a booking.

Gates are the main safety mechanism in a consumer agent, not a UX detail. An agent with broad permissions and no gate is a bad idea regardless of how good the model is.

When evaluating a product, check what sits behind a gate and whether you can adjust it. Purchases, outbound messages, and destructive changes should all qualify.

Sandbox (isolated session)

A sandbox is an isolated execution environment — commonly a per-user browser session — that keeps one person's credentials, cookies, and activity separate from everyone else's.

For an assistant that logs into your accounts, isolation is the difference between a private workspace and a shared one. Ask whether sessions are per-user and whether stored tokens are encrypted.

Integration

An integration is an authorised connection between an assistant and an external account — a mailbox, a calendar, a notes app — that lets it read or act on your behalf.

Integrations are where most of an assistant's practical value comes from, and where most of the privacy risk sits. Both facts are true at once, which is why declining to connect an account is a reasonable choice rather than a failure to adopt.

Inbox triage

Inbox triage is the task of sorting a large volume of incoming messages into the few that genuinely need a person, usually with summaries and draft replies prepared for the rest.

It is one of the clearest wins for assistants because the work is high-volume, low-judgement, and easy to check. The failure mode is an assistant that hides something important, so most products surface what they filtered rather than deleting it silently.

Hallucination

A hallucination is output that a language model presents as fact but that is not grounded in any real source — a plausible-sounding invention.

In an agent the stakes change. A chatbot that hallucinates gives you a wrong answer you can check. An agent that hallucinates may report a booking that never happened.

This is why the honest failure message matters so much: the right behaviour when an agent cannot complete something is to say so plainly, not to summarise imagined success.

An agent you can actually text.

Cael runs in iMessage and Telegram. Hand it something small and reversible and see what comes back.

Start 7-day trial

$0 today · Card required · Cancel anytime