Build your first AI support agent

What an AI agent does

An AI agent is a free team member that responds to customer tickets, looks up information through your connected integrations, and escalates to a human when it isn't sure. You bring your own OpenRouter API key and pay the model provider directly per token. No seat fee. Run as many agents as you want.

Each agent has:

  • A name and personality
  • A model (we recommend Gemini 3.1 Flash Lite Preview, benchmarked at 99.5% routing reliability and ~$0.38 per 1,000 conversations)
  • Skills (which integrations it can use)
  • A security level
  • An optional auto-respond switch

Setup

  1. Go to Settings → Team → AI Agents and click + New Agent.
  2. Give it a name. "Support Agent" works fine to start.
  3. Land on the agent's configuration page.

Five tabs walk you through the setup:

Character

Name, catchphrase, personality sliders (formality, verbosity, technical depth, warmth, humor, proactivity), and the system prompt (the agent's "soul"). Pick a template, tweak from there.

Model

Paste your OpenRouter API key. Keys start with sk-or-. You pay OpenRouter directly per token, not us.

Click the Recommended card at the top to use Gemini 3.1 Flash Lite Preview. That's our current benchmarked best pick for cost vs latency. You can swap models any time.

A note on what to avoid: some popular-sounding models (including Mistral Small 2603) failed our tool-use routing tests on OpenRouter. The Recommended card always points at the current verified-best model, so when in doubt, pick that one.

Skills

Toggle which capabilities the agent has.

Built-in skills are always available:

  • Library: search and read your help center articles
  • Customer lookup: find a customer's history
  • Ticket actions: create, update, assign, close

Integration skills appear only when the matching integration is connected:

  • Sentry, GitHub (code reads), GitHub Issues (file bugs), BetterStack, Slack, Sanity, Knowledge URLs (Atlas web reads)

Security

Pick a security level: Strict, Balanced, or Permissive. Set max messages before automatic handoff to a human. Toggle Auto-respond on the agent to reply to real customer messages on assigned tickets. Keep Auto-respond OFF until you've tested in the next tab.

Test Chat

Send the agent test messages and watch it use real tools. This is where you check that it picks the right skills, answers accurately, and doesn't hallucinate. Iterate on the prompt and personality until the responses feel right.

Deploy

Once the test chat looks good:

  1. Turn on Auto-respond in the Security tab.
  2. Assign tickets to the agent. Manually or via routing rules.
  3. Watch the conversations. Reassign to a human any time.

What it costs

  • cStar: $0. AI agents don't count as billable seats.
  • OpenRouter: paid directly by you, per token. Typically under $1 per 1,000 conversations on the recommended model.

That's it. Free agents, your model bill.