Instructions are the standing brief. Who the agent is, the ground rules, the voice, and what it should confirm before it acts.

“You are the support agent for Northwind. Greet the person by name once you know it. Confirm the account email before discussing billing. Answer from the help center and the customer’s record, and say plainly when something needs a teammate. Keep replies to a few sentences.”

That is the level to write at. Procedures belong in playbooks, and each playbook already carries its own description of when to use it, so instructions stay short. You can still name one when you want it preferred in an ambiguous case: “For anything about money back, follow the Refund playbook.”

The AI model is the engine behind the agent. Pick from the models available in your account, such as GPT 5.6, Opus 5, or Gemini 3.7. Heavier models reason better on messy threads. Lighter ones answer faster and cost less credits per conversation. Temperature sits next to the model: lower keeps answers tight and repeatable, higher gives more variety.

Iterations are how many think-and-act steps the agent may take on a single turn. The default is 10, and you can set anywhere from 3 to 30. A lookup plus a reply is a few steps. A turn that searches the CRM, calls Stripe, and updates a Ticket wants more room. Raise it when you see the agent stop early on complex asks.

Test in Overview before you connect a number. Send a first message, continue the thread, and close the test conversation when you are done.

The agent works with what you attach.

System tools are built into Frontline. Web search, URL scrape, and LinkedIn search. Turn them on when the agent should look beyond your data.

Custom tools are API calls your account defines. A billing lookup, a shipment status, a license check. Create the tool once, then enable it on the agent.

Data access covers People, Companies, Deals, Tickets, custom objects, and tables. You choose read, create, update, delete, and activity logging per record type. Support usually needs People and Tickets. Sales needs People, Companies, and Deals. An onboarding agent that files a document needs activity logging on People.

Integrations are connected apps: Stripe, a calendar, a help desk. Pick the account, and either all of its tools or a chosen subset.

Assigned knowledge bases become searchable sources, so help center pages and policy docs answer for themselves.

A playbook is a reusable procedure. It has a name, a description of when to use it, and the steps to follow. Assign it to the agent and the agent runs it when the moment matches.

The description is the match. “Use this when the customer asks for a refund.” “Use this when a Lead wants a demo.” The instructions are the work. Find the Person. Open the Deal or the Ticket. Do the steps. Say what happened.

This is how one agent covers many jobs. Support carries Refund, Billing question, and Password reset. Sales carries Qualify lead and Book a meeting. Collections carries Promise to pay. Onboarding carries Document request. Adding a playbook grows what the agent handles, with the core brief untouched.

Playbooks carry their own scope too. Stripe on the invoice playbook. Calendar on the booking playbook. People and Deals, read only, on qualify. The agent keeps a lean base and the library does the rest.

Write playbooks in Studio, or ask Max to draft one and assign it here. Active and assigned is what the agent can reach.

On WhatsApp the phone number arrives with the message. Frontline looks for a matching Person, and when there is no match it creates one, so the conversation is on a record from the first reply. The same happens with an email on live chat.

Which kind of Person it creates comes from the agent’s default record type. Set it to Lead on a sales agent and every new inbound number lands as a Lead. Set it to Contact on a support agent and known-customer traffic stays clean. This is one setting on the agent config, and it is the reason a qualifying agent can open with a real question instead of a form.

From there, details fill in as the conversation goes. A name from the WhatsApp profile, an email when they offer it, a company when they mention it. The agent writes them to the record it is already on, and matching on email keeps you at one Person.

Flows are graphs on the agent: say something, respond, run tools, call an API, branch, log an activity.

Since identity and contact details are captured for you, the best use of a flow is a stretch of work that has to happen in a fixed order, mid-conversation. Collect a policy number, then call the API, then branch on success or failure, then say an approved line word for word. Compliance wording, a payment link that depends on a lookup, a three-field intake for a specific product.

Every agent has a Start flow that runs at the first message (only available for live chat). Keep it light. A greeting reads better coming from the agent’s instructions, where it can react to what the person actually opened with, and the record already exists without a capture step. Reach for Start when you need exact wording on first contact, such as a recording notice or a regulated disclosure.

Other flows begin with a trigger. Either intents, which are phrases you train, or agentic routing, a sentence describing when to enter: “Enter when they ask to cancel.”

For most conversations a playbook fits better, because people wander and the procedure still holds. Flows earn their place when the sequence itself is the requirement.

An API node maps a field from the response into a variable, so a returned order number becomes {order_id} and the next message can say it back. Respond and tools nodes capture from the conversation itself. Create the variable first, then capture into it.

Write {VARIABLE_NAME} in messages, prompts, instructions, API URLs and bodies, and routing conditions. The name matches your variable exactly.

Variables are what the agent remembers inside one conversation. Contact fields come built in. You add the rest: deal_stage, ticket_id, doc_url, amount_due.

Each one takes a description, and optionally a pattern, such as an email shape, so a captured value has to look right. Frontline flags a name that is already taken or reserved, so pick something like email_api when an API value should stay off the Person record.

Workflows keep their own variables. A workflow that runs after the chat reads what the conversation produced, then writes to People, Deals, or Tickets with its own steps.

Connect each surface in Studio → Channels. A WhatsApp number, an Instagram account, or a Messenger page answers with one agent. Live chat is the agent’s own widget.

WhatsApp. Assign the number to the agent and test it with the QR code. Edit the business profile: about, website, email, category. On the agent’s WhatsApp settings you control split messages and the character limit, and how conversations close: never, automatically after a set time, or dynamically. Add a closing message, either fixed text or an instruction the agent writes from, and keep auto-close under 24 hours when a closing message is on. Idle settings say how long counts as quiet and how many times that clock may reset. Debounce turns a burst of short messages into one turn. Document and image processing let people send files and photos. A manual response timeout, up to 24 hours, decides how long the agent waits while a human has the thread, and auto reply picks the conversation back up.

Live chat. The same split and close controls, plus feedback on hover or always visible with the reasons you define, document and image processing, and idle. The theme covers title, first message, placeholder, avatar, bubble and message colors, alignment, progress indicator, and vertical position.

Messenger and Instagram. Split messages, close behavior, closing message, idle, image processing, manual response timeout, and auto reply after a handoff. Assign the page or the Instagram account to the agent.

Idle on the channel is what makes Conversation idle workflows possible, so switch it on when collections or onboarding should nudge people who go quiet.

Open Studio → Channels → WhatsApp and click the connected integration, then Create WhatsApp Button.

Two widget types. Suggested messages shows chips such as “I have a question,” “Connect with sales,” and “Connect with support.” WhatsApp conversation opens the chat straight away.

Set the phone number, whether the widget opens by itself and after how many seconds, whether it plays a sound, and the pre-filled message, for example “Hi, I’m interested in…”. Choose the icon and background color and the position on the page. The preview shows the chips, the draft line, and the “Chat with us” button.

This is your inbound door, and the chip someone taps tells you what they came for.

On the agent, look for Share your live chat agent or add it to your website.

Embed on your website gives you a script tag:

<script src="https://app.getfrontline.ai/frontline_assistant.js" data-assistant-id="YOUR_AGENT_ID" defer></script>

Paste it into your site and the bubble picks up the theme you configured.

Share via URL gives you a link to send. Same agent, no site changes, good for a campaign page or a partner.

Conversations is the agent’s inbox. Filter by open or closed, channel, search text, positive or negative feedback, and date range. Each row shows the Person, a rolling summary, and the channel it came in on.

Open a thread to read the transcript. Audit reveals the machinery behind each turn: flows entered and exited, intents matched, tools called, variables captured, and errors. Trace a single message when you want the model and reasoning behind that exact reply.

Intervene hands you the thread. You reply as the team while the agent holds. When you are finished, hand it back and the agent carries on. With a manual response timeout set, the agent returns on its own after a quiet stretch, and auto reply lets it re-open with the customer.

Intervene when collections needs a human on a payment promise, or when an onboarding thread has stalled on a document.

Use the ••• menu on a conversation.

Share conversation passes the thread to a teammate.

Export as .csv or .xlsx downloads the transcript with role, message, date and time, contact, type, and status.

Blacklist contact stops that person from reaching the agent on that channel, and their next message is ignored. Keep it for abuse and for numbers that should stay out of support.

Analytics is the agent’s scoreboard. Choose a date range and you get conversations and messages per day, average messages and average credits per conversation, a breakdown by channel, a breakdown by model, and feedback counts with the reasons people picked.

Watch the channel split to see where volume actually lands, and watch negative feedback with reasons to know which playbook needs a rewrite.

The agent handles the conversation. Workflows pick up when the conversation changes state, and you select this agent on the trigger so it fires for the right traffic.

Conversation ended. A workflow can transcribe the thread, analyze an attached document, write a note on the Person, update the Ticket, enrich the Company, or run an evaluation on quality. Collections marks the Deal and logs the promise. Onboarding attaches the file and sets a field. Support closes the Ticket and scores the chat.

Conversation idle. The person went quiet, and idle is enabled on the channel. The workflow reviews the thread and uses the Send message step to write back inside the same conversation. “Still need the signed mandate?” “Want me to resend the onboarding link?” That is a follow-up in context, on the thread they already know.

Send message writes into an open conversation on WhatsApp, Instagram, or Messenger. Send WhatsApp message delivers a template to a phone number or a People record, which is what you want when there is no live thread.

This is the most common build. The channel is WhatsApp, the agent qualifies, and the CRM fills itself in as the chat happens.

Someone messages the number from a WhatsApp button on your site or an ad. Frontline matches the phone number or creates a Person, using the agent’s default record type, so the new inbound arrives as a Lead with a live conversation attached.

Instructions set the job and the manners. “You qualify inbound interest for Northwind. Ask one question at a time. Find out what they are trying to solve, the size of their team, and their timing. Offer a call once there is real intent. Keep replies short and warm.”

A Qualify inbound lead playbook does the work. Confirm the name and the company. Capture the need, the team size, and the timeline. Write them onto the Person and the Company. Create a Deal once the intent is clear, and set the stage. Offer the calendar link when they want to talk to someone.

Give that playbook read and write on People, Companies, and Deals, and the calendar integration for booking.

Variables hold the answers mid-chat: need, team_size, timeline. The playbook writes them to fields when it has enough to be useful.

Idle picks up the half-finished ones. A conversation idle workflow reviews the thread and sends one message in the same chat: “Still want me to line up a time?” Most qualification dies to silence, and this is where you get it back.

Conversation ended closes the loop. Enrich the Company, score the conversation, and notify the owner when the Lead qualified.

Intervene when a rep wants the hot one. They take the thread, close the meeting themselves, and hand it back.

Support. Instructions set the voice and the confirmation rules. Playbooks cover refund, billing, and password reset. Runs on WhatsApp and live chat with document processing on. Conversation ended updates the Ticket and scores the chat. Intervene when the customer is upset.

Collections. WhatsApp, with idle on. A Promise to pay playbook updates the Deal and logs a note. A conversation idle workflow reads the thread and sends a reminder in the same chat. Conversation ended writes the outcome onto the Person and the Deal. A human intervenes when the customer wants to negotiate.

Onboarding. WhatsApp or live chat with document processing on. A Document request playbook explains exactly what you need. Conversation ended runs file analysis, logs a note, and updates the record. Idle offers to take the file again.

Start with the brief, one model, and two playbooks. Connect one channel. Add a workflow on ended or idle when the job continues after the chat. That is an agent ready for real traffic.