Skip to main content
This page walks through every section of the agent wizard. If you’d rather just get something on a phone, see the quickstart.

​
1. Start a new agent

From the sidebar, click Agents → New agent. Choose:
  • Blank — start from scratch
  • Template — start from a preset (sales, support, scheduler, etc.) and customize
Both create a draft you can edit freely.

​
2. Instructions

The Opening row at the top sets how each conversation starts. Click its pencil to choose who speaks first and, when the agent does, whether it says exact words or writes its own greeting from a prompt. The same window holds pre-recorded greeting audio and the quick-action buttons shown in chat. Below it is the prompt: the source of truth for how the agent behaves. Write who the agent is and how it speaks, what it should accomplish and in what order, and the rules it must never break. A good prompt covers:
  • Who the agent is and its tone (warm, concise, professional)
  • The goals it should accomplish
  • The questions it should ask, in order
  • Hard rules: never quote pricing, never give medical or legal advice, redirect questions about competitors
  • A closing behavior (confirm next steps, end gracefully)
Instructions support markdown. Use bullet lists for steps, and keep imperative voice (“Ask for the customer’s name” > “You should ask for the customer’s name”). Audit checks the prompt for gaps and conflicts, and Improve this suggests edits.

​
3. Voice & Speech

Everything but language applies only to calls.

​
4. Knowledge

Click Attach knowledge base and pick one or more KBs. The agent will retrieve from these whenever a caller asks a factual question. If you haven’t built a KB yet, see Uploading documents. A knowledge base’s contents aren’t versioned: adding or removing sources changes what the published agent knows straight away, without a deploy.

​
5. Tools & Actions

Tools the agent can call mid-conversation, including transfers, hanging up and escalating to a human. Click Add tool and start from a template, then change any of its actions. See Tools for the full reference. Common picks:
  • Transfer the call — say a line, then cold-transfer (or swap in a warm transfer)
  • Escalate to a human — hand a chat or email to your helpdesk queue
  • End the conversation — say goodbye, then hang up
  • Run a workflow — invoke a saved or inline workflow with parameters
  • Knowledge base lookup — explicit query (the agent already does implicit retrieval, but explicit lookups are useful when you want a specific KB queried for a specific reason)
  • DTMF — send tones to navigate IVR menus

​
6. Capabilities

Built-in abilities you turn on with a click: camera and screen share, image understanding, time awareness, widgets, web search, reading a webpage, continuing on the phone, and forms. Tiles that work on every channel come first; the channel-specific ones below say where they work. A pencil on an enabled tile opens its settings. See Built-in tools.

​
7. Before & After

Workflows that run around the conversation rather than during it: a pre-conversation workflow and identity verifiers before it starts, a post-conversation workflow and evaluations after it ends. See Setting up conversation triggers.

​
8. Deployment

Where you go live.
1

Assign phone numbers

Pick from your provisioned numbers, or buy a new one.
2

Publish

Click Publish in the top-right. This freezes a new version and routes traffic to it.
3

Optional: deploy as SIP

For BYOC setups, click Deploy as SIP to expose a SIP endpoint instead of a Twilio number.

​
Iteration loop

Most teams iterate as follows:
  1. Test in the simulator after every prompt change (Testing agents)
  2. Place a few real calls from your phone after each publish
  3. Review transcripts in Conversations, look for failure patterns, refine prompts and rules
Don’t try to make the agent perfect on day one — ship something safe, learn from real calls, iterate.