How to build and deploy your first AI agent in 15 minutes

Most AI support bots are frustrating. They give generic responses, can't actually solve problems, and end up creating more work for human agents. But it doesn't have to be this way.

In this guide, we'll build a support agent that can actually resolve customer issues. It will understand context, take actions in your systems, and know when to escalate to humans. Total time: about 15 minutes.

Step 1: Define the agent's scope

Before building anything, get clear on what your agent should handle. For this tutorial, we'll build an agent that can: reset passwords, check order status, process refunds under $50, and answer FAQs. Everything else gets escalated.

Start by creating a new agent in your dashboard. Give it a name like "Support Agent v1" and select the Support template as your starting point.

Step 2: Connect your tools

An agent is only as useful as the actions it can take. Head to the Integrations tab and connect: your customer database (we support Postgres, MySQL, MongoDB, and most CRMs), your order management system, and your knowledge base or help docs.

For each integration, you'll define what actions the agent can take. For example, for your database, you might allow "read customer info" and "update password" but not "delete account."

Step 3: Configure the LLM

In the Model tab, select your preferred LLM. We recommend Claude 3.5 Sonnet for support use cases — it's fast, accurate, and great at following instructions. Set the system prompt to define your agent's personality and boundaries.

A good system prompt includes: the agent's role, what it can and can't do, how it should handle edge cases, and when to escalate. We provide templates, but customize them for your brand voice.

Step 4: Build the workflow

Open the Visual Builder and you'll see a basic support flow. The default template includes: greeting, intent classification, action execution, and resolution confirmation. Customize each node for your use case.

For example, click on the "Intent Classification" node and add your specific intents: password_reset, order_status, refund_request, faq, and escalate. The agent will automatically route conversations based on detected intent.

Step 5: Test in sandbox

Before going live, test extensively. Use the Sandbox tab to simulate real conversations. Try edge cases: what happens if a customer asks for a $500 refund? What if they're angry? What if they ask something completely off-topic?

The sandbox shows you exactly what the agent is thinking at each step. You can see which intent it detected, what tools it called, and why it made each decision. Use this to refine your prompts and logic.

Step 6: Deploy

When you're confident, hit Deploy. Your agent is now live. You can embed it on your website, connect it to your helpdesk, or integrate via API. All conversations are logged with full traces for review.

Start with a soft launch: route 10% of tickets to the agent and monitor closely. Check resolution rates, customer satisfaction, and escalation frequency. Iterate based on real data.

Within a week, most teams see 40-60% of simple tickets fully resolved by the agent. That's hours of human agent time saved every day.