An Agent is the thing you build: instructions, the Areas it draws on, the Connectors it may call, and a model. There is no wrapper object between an Agent and the site it answers on.
Agent
Instructions, Areas, Connectors, model.
Modality
What it is to the outside — at most one.
Delivery
Where that appears — as many as you need.
Conversations
What people actually asked it.
Build the Agent
- Configure an Agent — purpose, instructions, and model.
- Use Areas — the containers that carry Knowledge, Standards, and Connectors.
- Always-on and on-demand Areas — what every conversation needs, and what should enter only when relevant.
- Attach Knowledge and Connectors — the smallest useful governed scope.
- Test an Agent — before anyone outside sees it.
Put it in front of people
- The website widget — the embeddable Delivery of an Agent.
- The AI overview on your search page — a search Agent's answer above your own results.
Operate with evidence
- Conversations and verdicts — read what visitors asked.
- Measure an Agent — the metrics, and when to distrust them.
Modality, Delivery, and the doors that are always open
Modality is what the Agent is to the outside: a chatbot, a search box, and later a voice agent or a guide. An Agent carries at most one, because the modality forks its instructions, its model, its tools and its settings.
Delivery is where that modality appears: a widget per domain, and later a channel. An Agent carries as many as it needs, freely, because a delivery forks nothing about the answer — only its appearance, its copy, its allowed domains, and what a visitor's browser may record.
The inner doors are neither. Chat, the API, MCP, and being called as another Agent's tool are open to anyone who may already read the Agent. They are not a card, a toggle, or a column, so an Agent with no Modality at all is finished rather than unfinished: it still answers on every one of them.