Agents
The agent editor and lifecycle, from drafting and saving to validating, testing, and publishing.
An agent is one language model, one system prompt, and a flat list of tools it can call during a call. The prompt describes who it is, the tools give it things it can do, and the model decides turn by turn what to say and which tool to reach for.
A workspace can hold many agents. Each has its own prompt, tools, voice, and phone numbers, and you build each one in the agent editor.
Editor layout
The Build tab is a document: Name and description at the top, then Greeting and Instructions. The capability rail sits beside it, or below it on smaller screens. Choose Model, Voice, Transcription, Background sound, Tools, Knowledge base, Caller memory, Variables, Call handling, Safety & privacy or Call checks to open a drawer. Closing the drawer returns you to the rail.
The shared header holds versions, the recorded reply time on a voice agent, Health, Test, publishing and save status. On a phone the reply time leaves the header; Overview still shows it. Health opens readiness checks and links to each fix. Builder sits at the end of the tab row. Automations, Tests, Channels and Settings have their own tabs alongside Build and Overview. Calls, Contacts and Dashboards live in the workspace navigation. Editor sections covers what every section sets.
Saving
Text editors autosave after you stop typing. Cards with a Save button need that explicit save. Pending autosaves are also flushed when you:
- Press Cmd/Ctrl + S.
- Start a test chat or call.
- Close or switch a capability drawer.
Saving is not the same as going live
Version-scoped edits stay in the draft until you publish. Agent-wide settings, including name, description, answering hours and call checks, apply as soon as they are saved.
Drafts and versions
Every agent has versions. At most one version is published at a time: the live version callers actually reach. Version-scoped changes are saved to a draft, which is the working copy.
The version control sits in the shared header and shows live and draft state. Open it to switch versions. The publish action beside Builder offers Create draft, Open draft or Publish, depending on the selected version.
Publishing
Publishing promotes a draft to live.
- On a published version, the action is Create draft. A published version is read-only, so to change a live agent you create a fresh draft from it, make your edits, then publish again.
- On a draft, the action is Publish. Review the changes from the live version, then confirm publishing. Errors block publishing, and so does an empty prompt. Risks and failed tests are listed before you confirm, and a failed test needs a tick to say you've reviewed it. New calls use the new version straight away; calls already in progress finish on the old one.
Versions are immutable once published, so each publish is a clean snapshot you can point back to.
What the validator checks
Before a draft can publish, the validator runs. Errors block the publish; risks flag weak spots you can acknowledge and ship past. Validation lists what it looks at.
Builder
Choose Builder at the end of the tab row, or press Cmd/Ctrl + \, to
open the side panel. Describe what
you want in plain language and it edits the agent through the same controls
you would use by hand. Builder shares the panel with Test, and the panel
stays available as you move between the agent's tabs.
add a tool that looks up an order by reference numberThree modes
- Plan only reads the agent and proposes a plan. It makes no changes.
- Ask me first proposes each change and waits for you to allow it.
- Apply now applies changes immediately, except a change that reaches live calls straight away, deletes something, or starts voice test calls or more than 20 test runs. Those always wait for you.
What Builder reads
Each message, Builder reads the agent's draft as saved, its readiness checks, the tests whose latest run failed with the reason each failed, and calls from the last 14 days that failed or set off an alert. It also knows what you have open: a call, a test run or a tool. For a question about how the platform works, it searches these docs and links the page it answered from.
Approvals
An approval card lists each change that reaches live calls straight away, with the value it would set, and a run's estimated cost in A$ with what that covers. If the agent changes between the card and your Allow, so the change would now do more than the card showed, nothing is applied and Builder asks again. An approval still waiting when you leave is still waiting when you open that chat again. If the change applied but Builder's reply after it did not arrive, the panel says so, and Retry asks for the reply again.
Undo
Before each change, the draft is saved: once per Builder message, and once per
change made over the MCP, through /v1 or with vai push. Say "undo that" and
Builder puts the draft back the way it was before your last change and lists
what it changed back. Say it again to go back one more change. Your newest ten
are kept for each agent. Over the MCP or /v1, call undo_last_change, with
steps to go back more than one change; from a terminal, run vai undo.
Fixing a call
Every ended call has Fix with Builder in its header. It opens the agent with Builder already on the job: Builder reads the call, turns it into a text test, and runs that test to show the failure. It proposes the smallest fix, applies it once you allow it, then runs the same test again and tells you whether it passes.
A test run that failed has the same button. There Builder works on the test that failed: it reads the run, fixes the agent, and runs the test again. A run that passed has no button. The readiness issues under Health have one too, which asks Builder to fix that issue.
Secrets
Secrets never reach the model. When a change needs a secret value, Builder asks you to type it into a dialog; the value goes straight to encrypted storage. Chat history is saved per agent.
Testing before you ship
Choose Test in the header, or switch the side panel from Builder to Test, and talk to the agent before any caller does. The agent greets you the same way it would answer a real call.
The test panel is never blocked by the validator, so you can test a half-finished draft freely. For repeatable, graded testing across many scenarios, see Evals.
Sharing a demo link
Share link under Channels mints a public URL anyone can open without signing in. Visitors see the agent's name and description and can start a voice call. No phone number is needed.
The link serves the agent's published version, so edits reach the demo only after you publish. Revoking the link kills the URL; re-enabling mints a fresh slug, so any link that leaked stays dead.
Caps on a shared call
Calls from the link run under the Share link section's caps (3 minutes and 30 turns by default). The same caps apply to the website voice widget and the typed conversation. See share links for the full list.
Archiving and deleting
Archive and delete, under the agent's Settings tab, holds the two lifecycle actions:
- Archive is reversible. It drops the agent out of the main list, and you can restore it later. Archiving unassigns the agent's phone number. The agent stops taking new calls shortly afterwards, not the moment you confirm. See Agent lifecycle for what that gap is.
- Permanent delete removes the agent for good. Use archive unless you are certain.