Configure your agent
Shape what your agent says and does in the visual workflow editor — the conversation flow, nodes, functions, variables, and actions.
After you create an agent, Intrlume drops you into the workflow editor — a visual map of the conversation. This is where you refine what the agent says, the steps it moves through, and how it decides where to go next. Open it anytime from Agents → your agent.
The conversation flow
Every agent follows the same shape: a system layer, a start node, one or more main task nodes, and end nodes.
| Piece | On the canvas | What it does |
|---|---|---|
| System | Global Context (dashed box) | The agent's identity, persona, and always-on rules. Assembled for you; editable if you need to override it. |
| Start | Greeting & Routing | Opens the call, reads the caller's first reply, and routes to the right next step. |
| Main task | e.g. Lead Qualification | The working steps of the conversation — qualify, book, collect, answer. |
| End | e.g. End — Qualified Lead | Terminal outcomes. Each end node maps to a disposition you'll see in call logs. |
Routing happens through functions, not lines. Each connection is backed by a function the agent can call. The lines on the canvas are just a picture of those functions — edit routing in a node's Functions tab, not by drawing arrows.
Edit a node
Click any node to open the Node Properties panel on the right. It has four tabs.
Prompt
The node's instructions — its objective and how to handle the step (for example, "Qualify the prospect by determining who the training is for."). Click expand for a full-screen editor, and type @ to insert a variable.
The Global Context prompt is assembled automatically from your agent's settings and shown read-only. You can override it from its expand view, but only do so if you know what you're changing.
Want to sharpen a node's prompt? Paste it into a general-purpose AI tool like Claude or Gemini, refine the wording there, and bring it back — a fast way to tighten the objective and instructions.
Functions
Functions are how the conversation moves. Each function becomes an LLM tool call the agent can invoke when the moment is right.
| Field | What it means |
|---|---|
| Name (snake_case) | The function name, e.g. route_to_main |
| Description (when to call) | Tells the agent when to use it, e.g. "Route the caller to the main flow to explore opportunities." |
| Route to | The target node — or a decision with conditions for branching |
| Properties (data to collect) | Inputs the agent gathers before moving on, e.g. intent (string, required) |
Actions
Things the node does automatically, separate from what the LLM says:
- Pre-actions run on entering the node — most often Say a fixed line (e.g. the greeting).
- Post-actions run on leaving — most often End the conversation (used on end nodes).
Settings
- Context strategy — Append keeps the full conversation history, Reset starts the node with a fresh context.
- Respond immediately — whether the agent speaks as soon as it enters the node or waits for the caller.
Prompting tips
Node prompts drive a spoken conversation, which has different rules than writing for chat. A few principles that lift results:
- Keep replies short — one or two sentences per turn; long answers become a monologue the caller forgets.
- One question at a time — ask, confirm the answer, then move on.
- Write for the ear — no bold, lists, or links (the agent reads it aloud); use connectors like "first… then… finally…".
- Spell out numbers, dates, and money the way they're spoken (e.g. "January second", "two hundred rupees").
- Give a fallback for uncertainty — tell the agent what to say when it doesn't know, so it won't guess.
- Be specific, and skip long "never say…" banlists — a short positive instruction works better.
- Test on real calls and iterate — judge changes across several calls, not one, and listen end to end.
Deeper references: Vapi's voice prompting guide and the general Prompt Engineering Guide. Your agent's system prompt is generated for you — these tips are for refining the node-level instructions.
Variables
Use {variables} to personalize what the agent says. Open the left panel to manage them:
- Variables — values you set yourself (e.g. a campaign name).
- Customer Vars — fields that come from your audience (e.g.
cust_name,cust_phone, and any custom columns).
Type @ in any prompt to insert a variable token. At call time, Intrlume fills it in per contact — so "Hi {cust_name}" is personalized on every call.
Knowledge & tools
- Knowledge Base (left panel) — link documents so the agent can answer from your content during the call. See Knowledge Base.
- Tools — actions the agent can take mid-call (look up a record, book a slot). Attach them to the node that should use them, and pre-fill any fixed inputs.
Voice, language & behavior
Click the agent card (top-left, with the name and language) to adjust:
- Voice & pace — the synthesized voice and how fast it speaks.
- Language — a primary language plus an optional second language it can switch to.
- Background sound — optional ambient sound (e.g. an office) so calls feel more natural.
- Locale — country, currency, and timezone for this agent (how dates and money are spoken, and scheduling).
The Lock language toggle (left panel) stops the agent from switching languages mid-call.
Test it
Test Call — try a phone agent live; for web agents, use the web embed to test in the browser.
Make one change, test, and check the transcript in Call Logs before the next change — small iterations are easier to debug than big rewrites.