Build your first agent
Create an agent, teach it from your documents, website and Q&A pairs, check what it finds, and test it in chat before anyone else sees it.
4 min readAbout the Agent Builder →
What you need first
- A workspace with an active plan — see Getting started.
- The material your agent should answer from: PDFs or Word files, your website address, and the questions customers ask most.
On this page
Create the agent#
Step 1: Open Agents and press New agent
The page shows how many agents your plan allows (for example "1 of 3 agents"). Choose Start from scratch — or Use a template to begin from a ready-made agent for your industry (see Templates).
Step 2: Basic Information
Give it an Agent Name (each name can be used once in your workspace) and a Description / Persona: who it is, who it talks to and what it should help with.
Step 3: Language Configuration
Choose the Default Language and the Tone — Professional, Friendly, Formal, Casual or Technical.
Step 4: Communication Style
Write the guidelines it should follow — what to say, what never to say. Press Save Agent.
Tip: You can change all of this later from the ⋯ menu on the agent's card: Edit Basic Info for the name and persona, Edit Style for the guidelines.
The agent's workspace opens. Its menu holds everything for this agent: Train Agent (knowledge), Workspace (workflows), Test Agent, Integration Code and the widget's look.
Depends on your plan: The number of agents depends on your plan. At the limit the button changes to Upgrade to add more. Compare plans.
Add knowledge#
Open Train Agent. Under Add a source there are four kinds of knowledge. Add as many as you need; each one is read and indexed in the background.
Step 1: Documents & text — files or typed text
- Upload a file: PDF, Word (.docx), plain text (.txt) or Markdown (.md), then Upload. Scanned PDFs are read too.
- Write it out: a Title and the Text, then Add Text Source — handy for policies, timings and prices that are not written down anywhere else.
Step 2: Link — your website
Enter the Website URL and press Add Website. The site is read in the background — the pages it links to and its sitemap — and the source shows Reading website, then Trained (or Failed, with Retry).
Tip: If your website shows prices, training on it can also fill your Products list, which the agent quotes from.
Step 3: Questions & Answer — your FAQ
Press Add FAQ and fill in a Source Name, the Question and the Answer. Each one saves a single question-and-answer pair.
Step 4: YouTube — videos with captions
Under Other ways, use Add YouTube Link with the address of one video. The agent learns from its English or Hindi captions, so a video without them cannot be used. Playlist and channel links are refused — add videos one at a time.
What this agent knows lists every source with its state — Indexed, Indexing, Queued, Failed or Stuck. Press Refresh to update it, and Retry on a source that failed or got stuck.
Good to know: A source cannot be edited. To change one, delete it (the bin icon) and add the new version. Spreadsheets are not used as knowledge — products and prices belong in Products.
Depends on your plan: The number of knowledge sources and the storage they use depend on your plan. Compare plans.
Check what it finds#
Before testing in chat, check that a question finds the right passage. On Train Agent, press Test Retrieval, type a question into Search Query and press Check Retrieval. Each result shows the matched text and its score.
Tip: If the right passage does not come up, the agent cannot answer that question well. Add the answer as a Q&A pair, or as typed text, in the words your customers use.
Test it in chat#
Step 1: Open Test Agent
From the agent's menu, or the Chat button on its card on Agents.
Step 2: Ask what your customers ask
Type, attach a file (PDF, photo, Excel or Word, up to 10 MB), or press the mic to speak. Ask in English, Hindi or Hinglish — the agent answers in the language the message was written in.
Step 3: Start over with New chat
History opens your past test chats.
Good to know: Until a workflow is published, the test chat says No published workflow — answering from knowledge only. The agent answers only when your content covers the question; when nothing matches closely enough, it says it could not confirm that instead of guessing.
Give it a workflow and go live#
Answering questions is the start. A workflow decides the rest — the welcome message, the menu, when to take a lead, write a quotation, book a slot or bring in a person. Build one in Workflows, press Publish, and then put the agent on your website or connect WhatsApp.
Once it is live, Conversations lists every chat with the agent, the channel and the customer; open one to read it. Your team can also talk to published agents from Ask Agent.
Related guides
Workflows
Lay out what your agent does on the builder canvas — menus, questions, lead capture, quotations, bookings, handoff, WhatsApp messages and conditions — then test and publish it.8 min readTemplates
Start from an agent made for your industry: find it, copy it into your workspace in one click, and make it yours.1 min readPut the agent on your website
Publish the agent, copy one script tag into your site, choose which websites may show it, and set its name and colours.2 min readFAQ and troubleshooting
Answers to the questions we hear most, and what to check when something does not work the way you expect.5 min read
Something here does not match what you see, or did not answer your question? Tell us — or look through the FAQ.