The agent editor

Tapping any agent — a card or member pill on the Agents page, or a node on the Builder canvas — slides its editor over the current screen. The drawer edge drags to resize, and the width is remembered.

Autosave: there is no Save button

Every change autosaves into the team's draft — the footer ticks through autosaving… to autosaved 12 s ago — and real callers stay on the live version until the team goes live. Opening the drawer never creates a draft; only actual edits do, so browsing is always safe.

Overview, model and voice

Overview holds the agent's name, description, type and team membership, plus Test in browser and Download config (the full configuration as JSON).

Model & voice (Model & behaviour for text agents) chooses the AI model from a searchable picker, with a temperature slider from Precise to Creative — see Models & platforms. For voice agents the same tab sets Language, Voice and Speech recognition; changing the language reloads the voices and recognisers that support it.

Conversation: greeting, context and prompt

The Conversation tab controls how calls open — No greeting, Say this exactly, or Generate from instructions — and holds the agent's full System prompt in a Markdown editor, the hand-edit surface for wording that matters precisely.

Three option cards sit between them. Call context picks which facts the agent is told at the start of every call — the date and time, the caller's number, the number dialled. Call length sets when calls are ended (unset, every call is capped at 5 minutes 5 seconds) and an optional time-up message. Silence check-in has the agent ask "Are you still there?" after a chosen stretch of quiet, optionally hanging up after three unanswered check-ins.

Tools & transfers

Three sub-tabs: Functions (custom tools, including REST calls with typed parameters), API keys (named secrets — write-only once saved), and MCP servers for external tool servers. Tools wired by the team graph appear read-only, labelled Managed by the team graph — edit in the builder. Each editable function has a Hide the result from the agent switch: the function still runs and its result is stored with the call, but the model only hears whether it succeeded.

Voice agents add the hand-back editor and two cards: Transfer behaviour — the comfort tone played during a transfer, the allowed-destinations pattern applied to every number the agent dials (with a live Try it tester), and the consultation brief for warm transfers — and Keypad input, which sets how the caller's key presses are gathered.

Deploy and Activity

Deploy lists everywhere the agent is answering. Each row's ⋯ menu offers Embed on your site (a copyable widget snippet for browser deployments), Move to another agent, and Undeploy. The Deploy to form targets the browser, an owned number or a SIP registration; the tab is also home to the telephone test and Place an outbound call — see Outbound calling.

For voice agents an Operations section closes the tab. If the model can't be reached sets a fallback chain, tried in order: restart the session on another agent, retry with another model, or transfer to a person. Call recording toggles recording for this agent's calls — recordings are encrypted at rest, and can optionally use a customer-supplied key, in which case the platform cannot decrypt them and in-dashboard playback stops working.

Activity shows the agent's most recent calls; the full history lives on the Calls page.

Hand-edit or ask the Builder?

Both write to the same draft. The drawer suits precise, known edits — this greeting text, that voice. The Builder chat suits structural change — new agents, rewiring, new tools.