Skip to main content

Build Your First Agent

This walkthrough builds a simple, single-skill agent so you can see the full loop — define, build, test — before tackling something more complex.

Prerequisite

Sign up for Automation Anywhere's free Cloud Community Edition (or use an existing tenant) so you have access to AI Agent Studio.

1. Define the goal​

Open AI Agent Studio and create a new agent. Give it a narrow, concrete goal to start — broad goals are harder to test and harder to trust. A good first goal: "Given a customer's product complaint, recommend whether to offer a refund, a replacement, or an escalation to a specialist."

Write this goal down in plain language before touching any configuration. Everything else you build should trace back to it.

2. Create a skill​

Add a new skill to the agent, for example recommend_resolution. A skill needs:

  • A prompt — the instructions describing exactly how to complete this task, including the categories it should choose between (refund / replacement / escalate) and the criteria for each.
  • Inputs — the information the skill needs to do its job (in this case, the complaint text and any order details).
  • An output format — a structured result the rest of the agent (or a downstream system) can act on, rather than free-form text.

Keep the first version of the prompt simple. You'll refine it after you see how it performs on real examples.

3. Connect data (optional, for a first pass)​

If you want the agent to ground its recommendation in real order history rather than just the complaint text, connect a data source (for example, a CRM or order database) so the skill can look up relevant records. For your very first agent, it's fine to skip this and test with the complaint text alone — you can add grounding once the basic flow works.

4. Add a guardrail​

Add at least one guardrail before testing: for example, require human approval before any recommendation that results in a refund above a certain dollar amount. This is a good habit to build early, even on a test agent — it's the same pattern you'll rely on for production agents.

5. Test the agent​

Use AI Agent Studio's testing tools to run the agent against a handful of sample complaints — including edge cases (an angry customer, a vague complaint, a complaint with no order number). Check:

  • Did it choose a reasonable category for each case?
  • Did the guardrail trigger when it should have?
  • Where did the reasoning go wrong, and does the prompt need to be more specific?

6. Iterate​

Refine the skill's prompt based on what you saw in testing, and re-run the same test cases. This test-and-refine loop is the core workflow of building agents — expect to go through it several times even for a simple agent.

What's next​

Once this single-skill agent works reliably, the natural next steps are: adding a second skill (so the agent can handle more than one type of request), connecting a real data source for grounding, and deploying the agent so it can be invoked from Control Room or called by other agents.