Agents

Each workspace has one agent. The agent editor is where you build it: a sectioned form, not a canvas. The left rail groups every section; this guide maps them and explains the two things that trip people up — saving and publishing.

The sections

The editor rail is grouped by what you are doing.

Agent

• Overview — status, readiness, and recent calls at a glance. The agent's landing page.
• General — name, description, and the timezone the agent reads the clock from on a call.

Design

• Prompt — the persona and rules. The model reads it on every turn.
• Tools — what the agent can call during a call. Default and per-tool filler phrases live here too.
• Variables — tokens substituted into tool URLs and headers at call time, plus secrets.

Voice

• Voice — how the agent sounds, and its greeting.
• Model — the language model powering the agent.
• Conversation — how the agent speaks and behaves on a call: speed, tone, response length, ambience, and fillers. Covered in Voice and fillers.

Deploy

• Phone — inbound numbers routed to this agent.
• Share — a public demo link anyone can open, rate-limited and time-capped.

Compliance

• Outcomes — labels a call can be tagged with when it ends. Drives analytics grouping.
• Privacy — per-agent retention and PII redaction, overriding the workspace defaults.

Danger zone

• Archive — archive the agent (reversible) or delete it permanently.

Saving

Most edits autosave as you make them — there is nothing to click. The Prompt page and a few others save the moment you stop typing. Before you open the test panel or publish, the editor flushes any in-flight edits so what you test is what you have written.

The Builder Chat

The right rail is the AI editor — Builder Chat. Describe what you want in plain language ("add a tool that looks up an order by reference number") and it edits the agent for you against the same controls you would use by hand. It runs in three modes:

• Plan — reads and proposes, makes no changes.
• Review — proposes each change and waits for you to approve it.
• Auto — applies changes immediately.

Secrets never reach the model: when a change needs a secret value, the editor asks you to type it into a passive dialog instead of sending it through the chat.

Publishing

Your edits live in a draft. Publishing promotes the draft to the live configuration. Before it goes live, the validator runs the same checks the editor shows as red dots on the rail. Issues come in two severities:

• Errors block publishing. The publish is refused and the blocking issues are listed so you can jump straight to each one and fix it.
• Risks do not block. They are surfaced so you can decide whether to address them or acknowledge and ship anyway.

A published version is read-only. To change a live agent, create a fresh draft, make your edits, and publish again. Only one version is published per agent at a time.

Testing before you ship

Open the test panel and talk to the agent before any caller does — desktop has a side rail, mobile has a chat button. The test panel is never blocked by the validator; only publishing is. Iterate on the prompt and tools until the agent behaves correctly every call, not just most of them. Reliability is the bar. For repeatable, graded testing see Evals.