Concepts
Updated 2026-09-16
Eight words are enough to understand everything else.
Agent
The identity: name, persona, engine and model. An agent is yours: it lives in your account, and from there it is activated on whichever channels you want it to work in. The same agent can answer on WhatsApp, in a Teams workspace and in your editor without being duplicated.
Machine
Each agent runs in its own isolated machine: its own disk, its own terminal, no neighbors. That is where it reads your files, runs code and stores what it learns.
- It sleeps on its own when nobody is talking to it and wakes on the first message (a few seconds). You don't pay for time asleep.
- It persists: what the agent wrote to its disk is still there when it wakes up.
- It recovers: if the machine is lost, Studio recreates it with the account's history.
Disk limits and how many machines can be awake at once depend on the .
Conversation and turn
A conversation (session) is a thread with memory: the agent remembers what was said in it. A turn is what happens between your message and its final reply.
A turn is not an HTTP request: the work lives on the server. You send the message, get 202 Accepted right away, and follow progress through an you can connect to, disconnect from and come back to without losing anything. A turn can last seconds or an hour, and you can .
Within a turn the agent may ask for permission for a sensitive action. If your client knows how to render that question, it answers it; if not, the turn runs in automatic mode.
Channel
Where the message comes in from: , , the web chat, an , an or . The agent is the same; what changes is the context it is given and which tools it is allowed to use (fewer on a public channel).
Connector
An external account the agent can access on your behalf: Google, Canva, Spotify, Odoo… Connectors belong to the account, not to the agent: you connect Google once and any of your agents can use it. See .
Skill
A packaged instruction that teaches the agent to do a complete task — produce a designed PDF, clone the layout of a Word file, proofread spelling — with the tools it already has. Skills live in the agent's machine and are read when the task calls for them.
Engine and model
The engine is the harness that runs the turns; the model is the brain that answers inside it.
You can use your own provider keys: they are stored encrypted in your account's vault and only injected into your agent's machine.
Plan and usage allowance
Each plan includes a monthly token allowance. When it runs out, new turns respond 402 with the reason and the agent does not start work it won't be able to finish. See and .