Terms

The words are laid out by what rests on what. Nothing here needs learning in advance; come back when a word starts getting in the way.

An agent and what it is made of

Agent is a standing entity: its own model, its own instructions, its own access, its own channel. It exists whether or not anyone is talking to it right now. You can write to it, and it can start itself on an event.

Skill is a description of how an agent should do its work. A skill describes the work; it does not widen access. Very often a skill needs tools, and those come from a connector.

Model, in our context, is an LLM: a neural network trained on a body of data, a program that generates content. The model is what receives the user's messages and decides what to answer or which tool to reach for. Access to models comes from LLM providers.

Token is the unit a model measures text in: both what it was sent and what it answered. Spend is counted per run and per model and shows up in the journal. You pay your own LLM provider at your own rate. The word has nothing to do with keys or session tokens.

How an agent reaches the outside world

Connector is the one point where the platform extends and agents gain new abilities. One integration brings tools, triggers, prompt blocks and repeatable jobs. The Mail connector, for instance.

Connection is an instance of a connector holding your credentials. The connector describes what the integration can do; the connection says on whose behalf. A connection to one particular mailbox, for instance.

Tool is a command or an instruction on a connector that an agent can call. Send a message, read a sheet, file a task. One connection can carry many tools.

Trigger is an event that starts an agent: an incoming message, a deadline that came up, a change in a service. Triggers are what makes every agent proactive. They are what puts agents on a par with self-directed colleagues, able to automate your routine and even do the work for you.

Channel is what an agent interacts with you through. It can be the built-in chat or a Telegram bot connected through a connector.

App is a companion program on your device, registered with a key: a desktop client, a phone, a browser. Once linked, it brings the agent its own tools and triggers. The agent still has no direct access to the machine, and everything is bounded by what the app can do.

Who is allowed what

Credentials of a connection are stored encrypted on the platform and never reach the model: the agent names a tool, and the key is filled in outside the model. An LLM provider's key is what lets the platform address the model itself and run the agent on it.

Policy is an allow or a deny on one tool or trigger inside a connection. It is checked before the external service is called, or before the agent is.

A connection you have not opened to a particular agent does not exist for that agent. An open one is allowed whole, with all of its tools and triggers, until a policy forbids the excess.

What is left behind after the work

Run is one handling of one event: from the trigger or the message through to the answer. Step is one stage inside a run: a call to the model, or a tool call with its parameters and its result.

Session is a conversation with an agent, whatever carries it: a chat tab, a messenger, an IDE. Usually one session is one topic of conversation with the agent.

Journal is the runs, steps and calls you can open and read: what the agent did, with which parameters, and what the service answered.

When there is more than one agent

Agentic team is several agents over a shared task board, where it is visible who took what.