Skip to main content

Build and Test Your First Agent in the Console

In about thirty minutes you can go from a blank workspace to a working agent that answers real questions, such as "When is my policy renewal due?" You do not need to write code. Everything happens in the Achilles web console.

Before you start​

You need three things:

  • A sign-in to your Achilles console, from your workspace admin.
  • A short process document (an SOP) that describes what the agent should do. A one-page draft is fine. See Prepare your SOP.
  • Optionally, a connection to a data source, so the agent can look up real records. Your IT team can help. See Connect your data.
tip

No data connection yet? You can still build and test the agent. It will answer from your SOP, and you can add data later.

Step 1. Sign in and pick your workspace​

  1. Open the console link your admin sent you and sign in.
  2. Choose your workspace. A workspace is your organisation's private space. Nobody outside it can see your agents or data.
  3. Choose a project. A project groups the agents for one purpose, for example "Motor renewals" or "Loan servicing".

If you only see one workspace and one project, you are already in the right place.

Step 2. Create an agent with Agent Builder​

  1. In the left menu, under Build, open Agent Builder.
  2. Give your project a clear name.
  3. Upload your SOP, if you have one. PDF, Word, Markdown and plain text files all work.
  4. Answer the interview questions. The builder asks about your goal, your customers, what the agent may look up, what it may do, and when it should pass a customer to a person.
  5. Review the Blueprint, which is a plain summary of everything you told the builder. Approve it, or write feedback and let the builder revise it.
  6. Choose Build. The builder creates the agent and checks it for problems.

You can close the tab at any point and come back later. Your progress is saved. Learn more in Agent Builder.

Step 3. Test it in Playground​

  1. Open Playground in the left menu and pick your new agent.
  2. Ask questions the way a real customer would. Try "Show me my renewals this week" or "What is the status of my claim?"
  3. Watch what the agent does. Playground shows which lookups it ran and what came back.
  4. Try to break it. Ask for something it should refuse, such as another customer's data. Ask something off topic. Say you want to speak to a person.

If an answer is wrong, edit the agent, or update your SOP and rebuild, then test again. More tips in Test your agent.

Step 4. Check the health score​

Open Evals and Health. The health score tells you whether the agent is well set up and safe to launch. Fix anything flagged as blocking before you go live.

What "go live" means​

Until you go live, only people inside the console can talk to your agent. Going live means two things:

StepWhat it does
DeploySaves the agent and makes it active in your workspace.
Publish to a channelPuts it in front of customers or staff, for example on your website or in WhatsApp.

You can deploy without publishing. That is a good way to let a small internal group try it first.

Go live when your test conversations look right and the health score has no blocking issues. See Website widget and Messaging channels.

note

Your agent follows safety rules that are enforced by the platform, not left to the AI's judgement. See Safety overview.

Where to go next​