Core concepts
Six ideas the whole platform is built on.
Each of these is a deliberate design decision with a call-quality argument behind it.
Workspace
A workspace is your tenancy: one team, one billing account, one set of resources. Phone numbers, members and billing belong to the workspace rather than to any single agent.
Numbers sit at the workspace level so they cannot be stranded. An agent can be archived or deleted without a live number going with it.
Agent
An agent is the thing that answers the phone: one language model, one system prompt, and a flat list of tools. A workspace can hold as many as you need, each with its own voice, knowledge base and assigned numbers.
There is no diagram and no graph to wire up. A branching graph encodes the call you imagined. Real callers answer two questions in one breath, change their mind, or ask something no branch anticipated. Every one of those is a dead end in a graph and an ordinary turn for a model that reads a prompt and picks a tool.
System prompt
The prompt is the persona and the rules: who the agent is, how it speaks, what it must and must not do. The model reads it on every turn.
Every new agent starts from a starter prompt written in sections, because an agent follows a sectioned prompt far more reliably than a wall of text.
Tools
Tools are how the agent does real work: look up an order, book a meeting, transfer to a human, end the call. Each tool carries a description the model reads to decide when to call it.
If an agent calls a tool at the wrong moment, or never calls it, fix the description before you touch the prompt.
A tool can name prerequisites that must succeed first. Until they do, that tool is hidden from the model entirely. You get sequencing from a list of names, not a diagram.
Versions and publishing
Your edits land in a draft. Publishing promotes that draft to the live version that answers real calls. A published version is read-only, so changing a live agent means creating a fresh draft, editing it, and publishing again.
Publishing runs the validator first. Errors block, risks warn. The same checks show in the editor as you work.
Not every setting belongs to a version. Some take effect on the next call with no publish step. The version boundary has the full split.
Calls
A call is one conversation between a caller and the agent. Every call is transcribed, and the transcript flows into calls and analytics.
Scoring is opt-in. A metric runs only once you enable it, and an agent with none enabled is scored by nothing and charged for nothing.