Setting Up Your Agent

Creating an agent#

Two ways in, both from Agent → Agents:

1. 1Start from a template - Payment reminder, Appointment confirmation, Lead qualification, or Support follow-up. Each arrives as a complete brief you edit rather than a blank page.
2. 2Or click Create agent for a blank one.
3. 3Either way a chat opens: describe the agent in your own words ('chases overdue invoices for a Mumbai logistics firm, polite but firm, never threatens legal action') and it drafts the identity and the instruction document for you.
4. 4Keep talking to refine it, then open the agent to review everything in the three setup steps below.

1 · Identity#

Who the agent is and how it sounds. These feed every reply on every channel.

2 · Instructions#

This is the agent's brain, and it is written per channel, not once for everything. Voice, WhatsApp, Instagram, Email and Widget each get their own document, because the same instruction that works in an email is wrong out loud. Switch channels with the control above the editor.

Each document starts from a channel-shaped skeleton - WhatsApp's looks like this:

## Persona
-

## Environment & Situation
- Channel: WhatsApp chat. {{conversationType}} is inbound or outbound for this thread.

## Objective
-

## Writing Style
- Short chat bubbles. One idea per message. No sign-off.

## Facts
-

## Conversation Guidelines

### Phase 1:
-

## Guardrails
- 

Greetings: inbound and outbound differ

Each channel has two opening lines. The outbound greeting is used when you reached them (a campaign call, a first-touch email). The inbound greeting when they reached you. Leave inbound empty and it falls back to the outbound one.

Variables#

Anything in {{double braces}} is filled in at send time. Built-ins cover the conversation itself - {{userName}}, {{agentName}}, {{merchantName}}, {{conversationType}} (inbound or outbound), plus {{current_date}} and {{current_time}} in your timezone.

You can define your own too, each with a label and an example value, and fill them from a campaign CSV - {{invoiceNumber}}, {{dueAmount}}. A variable with no value at send time renders as (not provided) rather than being silently dropped, so a missing field is visible instead of producing a confident wrong sentence.

3 · Knowledge & rules#

• Rules - short, absolute lines the agent follows on every channel ('never quote below the list price', 'always offer a callback before ending'). Keep them few and unambiguous. They carry more weight than prose.
• Sample conversations - real exchanges showing how you'd handle a situation. Tone copies faster from examples than from adjectives.
• Knowledge tables - which of the workspace's knowledge sources this agent may read. Attach only what's relevant. A support agent doesn't need the sales deck.

Knowledge itself (uploads, URLs to crawl, FAQs) is covered in the Knowledge Base page - this step is about which of it this particular agent uses.

Autopilot & channels#

The right-hand sidebar is the live control panel. The switch at the top is the master: turning it off pauses the agent and clears every channel auto-reply at once, so incoming messages stop getting automatic responses until you turn it back on.

Below it, each channel is a row: pick which workspace connection this agent uses, and whether it may auto-reply there.

Voice has one extra setting: action mode. In draft mode the actions a call produces (an email with details, a callback invite) wait in the inbox for approval. In auto they go out during the call.

Auto-send confidence#

Every drafted reply gets a confidence score. Above the threshold it sends by itself. Below it, it waits in the inbox as a draft. The slider defaults to 90%.

• Auto-send needs both: the score must clear the threshold AND that channel's auto-reply toggle must be on.
• 70% is the floor. Dragging below it asks you to confirm, because that's where replies the agent isn't sure about start going out unreviewed.
• The threshold is per agent, so a support agent can be looser than one that quotes prices.
• The slider is disabled while the agent is paused.

Follow-ups#

Each agent decides when silence deserves a nudge. Defaults and units differ by channel because the platforms differ:

Full detail, including what happens when a free-form message is rejected outside the window, is on the Follow-up Automation page.

Multiple agents#

Run as many agents as you need - one per product line, market, or job to be done. There is no per-plan agent cap. Agents consume credits only when they actually work.

• Each agent has its own instructions, channel bindings, knowledge tables, custom tools, follow-up schedule and confidence threshold.
• A channel connection belongs to one agent at a time - two agents can't answer the same WhatsApp number.
• Members can be granted or denied access per agent, so a junior rep sees only the agent they work on.
• Switch agents from the agent selector. The inbox, analytics and setup all follow the selection.