Skip to main content
Explore Agents panel in light mode
Agents are internal AI assistants your team uses inside the platform. They search transcripts, summarise conversations, run reports, draft flows, schedule background work, and generally let you point natural language at the product instead of clicking through menus. They never face customers — for customer-facing AI, build a Flow.

​
Where agents live

Internal agents are not on the Flows page anymore — they’re surfaced through four fixed buttons at the bottom-right of every page:
ButtonWhat it opens
Chat with AgentThe chat panel for whichever agent you last used.
Chat HistoryPast chat threads, searchable and resumable.
Explore AgentsThe list of every agent in your tenant, plus the + button to create a new one.
Scheduled Agent TasksBackground jobs you’ve handed off to agents (e.g., “every Monday, summarise last week’s flagged transcripts”).
The Explore Agents panel slides out from the right and shows each agent with its name, one-line description, and a kebab menu (⋮) for edit / delete actions.

​
Talking to an agent

Click any agent in the Explore Agents panel to open the chat panel pre-targeted at that agent. Or click Chat with Agent (which resumes the last-used agent). Send a message; the agent uses its configured prompt and functions to respond.
Agent chat panel in light mode
The chat panel supports text, voice (speech-to-text), and file attachments. Replies render Markdown — headings, bullets, tables, links.

​
Creating an agent

Click the + button at the top of the Explore Agents panel (or navigate to /Agents/New). The editor opens to the Prompt tab.

​
Configuring an agent

Agent Prompt tab in light mode
Each agent has five tabs.

​
Prompt

The system instructions that define the agent’s role, capabilities, and behavior. Includes:
  • Name — shown in Explore Agents and the breadcrumb.
  • Description — one-line summary, also shown in Explore Agents.
  • Model Type — pick the model variant. Fast (lower latency, lighter reasoning) vs reasoning-tier (slower, deeper).
  • Prompt — rich-text editor for the agent’s instructions. Supports headings, bullets, bold/italic.
  • Suggested Messages — clickable starter prompts shown when a user opens a fresh chat. Each suggestion has a label (what the user sees) and a prompt (what gets sent).
A workable structure for the prompt: 
Act as a <ROLE> within our AI contact center platform. You will help internal staff <PURPOSE>.

Workflow & Capabilities
- <capability 1>
- <capability 2>
- <capability 3>

Steps for <COMMON TASK>
1. Call <toolA> to gather…
2. Call <toolB> to filter…
3. Return a short summary.

Formatting
- Use Markdown headings and bullets
- Keep responses brief
- Never return raw data dumps unless requested
Use direct verbs (“must”, “will”) instead of suggestions (“should”, “can”). Agents follow declarative instructions more reliably than aspirational ones.

​
Functions

Agent Functions tab in light mode
Pick which JS functions the agent can invoke. Click Add Functions to open the picker, then select the functions to grant. Each enabled function becomes a tool the agent can call during a conversation. Common picks for a Conversation Analyst Agent:
  • listTranscripts, countTranscripts — search and count.
  • getTranscriptReport, getReadableTranscript — inspect.
  • getTagsForTranscript, setTagsForTranscript — read and write tags.
  • searchDocuments — ground in playbooks and SOPs.
The fewer functions you enable, the more predictable the agent’s behaviour. Resist the urge to grant every function “just in case” — every extra function adds reasoning load.

​
Memory

Long-lived facts the agent should always know. Add entries for things like “the company’s fiscal year ends in March”, “our SLA is 24 business hours”, “we never discount more than 10% without manager approval”. Memory entries persist across every chat and apply to every conversation the agent has.

​
Variables

Static parameter values that the enabled functions require. When you add a function that needs configuration (an API key, a tenant ID, a base URL), that requirement surfaces here so you can supply the value once instead of per-call.

​
Branding

PDF and PowerPoint export styling for agents that produce documents. Set a company name, primary/secondary colours, upload a logo (PNG/JPEG/SVG ≤ 2 MB), and optionally write free-form logo-usage instructions (“logo top-right, page 1 only”). Leave blank to use platform defaults.

​
Scheduled agent tasks

The Scheduled Agent Tasks button (bottom-right of every page) opens the schedule manager. Set up recurring or one-off jobs an agent should run on its own — e.g., “every weekday at 8am, list yesterday’s flagged conversations and post a summary to a thread”. Useful for handing off the kind of routine analysis that doesn’t need a human in the loop.

​
Use cases

Agent purposeFunctions typically enabledNotes
Conversation AnalystlistTranscripts, countTranscripts, getTranscriptReport, getTagsForTranscript, setTagsForTranscriptQA and reporting.
Flow BuildersearchFlows, createFlow, createState, editState, createTransitionDrafts flows from a natural-language brief.
Campaign ReporterlistOutboundCampaigns, getCampaignStats, searchTranscriptsPull-the-numbers reports on outbound results.
Knowledge CuratorsearchDocuments, listFaqs, createFaqFind gaps in your knowledge base and propose new FAQs.

​
Agents vs Flows

AgentFlow
Who talks to itYour team, inside the platform.Your customers, on a channel.
SurfaceChat panel (bottom-right).Voice, SMS, email, web chat, etc. via a Point of Contact.
Logic shapeSingle prompt + tool list.Stateful graph with transitions, events, and per-state thoughts.
Published?Always live for everyone in your tenant.Has draft and published versions.
Testable howJust chat with it.Chat sandbox + Test Scenarios.
If you’re not sure which to build: would a customer ever talk to this? If yes, build a Flow. If only internal staff, build an Agent.

Functions and events

Authoring the JS functions agents call.

Flows

Customer-facing equivalent.

Documents

Ground your agents in real source-of-truth content.

Schedule

When the tenant is open — relevant for scheduled agent tasks.