Agent Builder

Agent Builder is the fastest way to create an agent. It works like a conversation with a business analyst. You describe your process and the builder turns your answers into a working, checked agent. You review and approve before anything goes live.

You find it in the console under Build, in the Agent Builder menu item.

The three stages

Stage	What happens	What you do
Interview	The builder asks questions one at a time.	Answer in your own words.
Blueprint	The builder shows a plain summary of what it plans to build.	Approve it or send feedback.
Build	The builder creates the agent and checks it.	Review the result, then test.

What you provide

You can start with very little. The more you give, the better the first draft.

• A project name. Pick something your team will recognise.
• Your SOP or process documents (optional). Upload PDF, Word, Markdown or text files. The builder reads them and pulls out the steps, roles and business rules. See Prepare your SOP.
• A data connection (optional). If the agent should look up policies, loans or claims, connect your database. The builder explores what is there and asks you which parts the agent may use. See Connect your data.
• Answers to the interview. Expect questions about your goal, your customers, who uses the agent, what it may look up, what actions it may take, and when it must hand over to a person.

note

The builder hides sensitive-looking details, such as ID numbers, before it reads sample records. They are never shown to the AI or stored in the session.

What you get

When the build finishes you have:

• A working agent with its instructions, abilities and roles set up. See Agent abilities and Roles and access.
• A set of starter test cases based on your SOP.
• A health score showing whether the agent is ready to launch.

Review and edit

Do not skip the blueprint. It is the cheapest place to fix mistakes.

1. Read each section. Does it match how your business actually works?
2. Check the roles. Are the right people allowed to do the right things?
3. Check what needs approval, especially anything that changes a record or sends a message.
4. If something is wrong, type your feedback in plain language. For example: "Renewal reminders should only go to customers with a valid mobile number." The builder revises the blueprint.
5. When it looks right, approve it.

After the build, you can keep refining. Open the agent in the console, adjust its instructions or abilities, and test again in Playground.

The deploy gate

The builder will not deploy an agent that has blocking problems. Before the agent goes live, the platform runs your test cases and calculates a health score. If something serious is wrong, for example a missing safety rule, you will see it flagged and the deploy button stays locked until you fix it.

Smaller issues appear as warnings. They lower the score but do not block you.

Tips for a strong first build

• Start with one narrow job, such as claim status. Add more later.
• Write the "what should never happen" list. It becomes safety rules.
• Include two or three real customer questions in your SOP.
• Save and resume freely. Your session is kept.

tip

If the first build is close but not perfect, refine your SOP and run the interview again. Changing the document is often quicker than fixing the details one by one.

Next steps

Try your agent in Test your agent. When you are happy, follow Website widget or Messaging channels to launch.