Build an agent with AI

Create with AI replaces the whole setup with an interview. You answer six questions, optionally hand over files and website addresses, and a model designs the agent's prompt, greeting, voice and model settings. The build continues after you close the dialog, so you can carry on working and pick the agent up on the agents list when it is ready.

What you need

• Membership of the workspace you are building in.
• At least one active voice in the library, since the agent's voice is picked from the voices on offer.
• Room in your plan's active-agent limit, which is checked when the agent is finally created.
• A microphone, only if you want to answer by voice instead of typing.

Steps

Step 1: Open the flow

On the agents list select Build New Agent, then Create with AI.

Step 2: Answer the six questions

They come in this order, one screen each:

1. What is the business, and who is this agent to a caller, its name and role?
2. Walk through a typical call, hello to goodbye. What do callers most often want?
3. Who calls, how should the agent sound to them, and in what language and dialect?
4. What must the agent never do or say?
5. When the agent can't help, or the caller asks for something out of scope, what should it do?
6. Is there anything it may need that isn't in your uploaded documents?

Question 1 offers the chips Retail, Clinic, Real estate, Restaurant, Education and Something else. Question 3 offers Egyptian, Gulf, Levantine, Fusha, English and Mixed. A chip fills the field with its own text, which you can then edit.

Enter sends the answer and moves on, and Shift with Enter starts a new line. Back returns to the previous question with your answer still there.

Step 3: Answer by voice, if you prefer

Select Answer by voice to record instead of typing. The clip is transcribed and lands in the field as editable text, so a mishearing stays fixable. Recording stops by itself after five minutes, and whatever was captured up to that point is still transcribed.

Step 4: Hand over what the agent should read

The last screen asks Anything we should read first? and is optional. Use Choose files to upload documents, and the address field to add websites. Each file and address becomes a knowledge source in your workspace library, so it is available to your other agents too.

Step 5: Start the build

Select Build my agent. The dialog switches to Building Your Agent and tells you this runs in the background. Close it.

Get it right

The six answers are the whole brief. Write each one as a paragraph with the specifics you would give a new hire, because the design can only be as precise as the answer it reads.

Recommended

We are Acme, an online electronics shop in Cairo. The agent is Sara, the first
person a customer reaches when they call about an order. Most calls are about
where a delivery is, how to return something, or getting back into an account.
Callers are Egyptian, mostly on a phone in a noisy place, so keep it short.

Less effective

Customer support for a shop.

Worked example

You answer the six questions for Acme, upload the returns policy as a PDF, and select Build my agent. A card appears on the agents list reading Reading your sources, then Designing your agent, then Almost there.

A minute or two later the card is replaced by an agent called Sara, with the Agent ready toast and the body {name} is built and ready to use. The returns policy is attached on its Knowledge tab, and you open Persona to check the prompt before you connect a channel.

What happens next

A card appears on your agents list with the Building badge, showing where the build has got to:

The card carries no link: there is no agent to open until the build finishes, at which point the card is replaced by the real agent row and you get the Agent ready toast with the body {name} is built and ready to use.

The name on the card while it builds is a placeholder, taken from the first line of your first answer and cut to 60 characters. The real name comes back in the design.

What the design decides, and what you do

The design picks the name, the description, the prompt, the greeting, one voice with its speech settings, the model and temperature, the retrieval settings, and which of the three built-in tools are on. Each field is checked before anything is written.

Three things are always yours:

• Knowledge. The agent is given the sources you added during this build.
• Time zone. A new agent is created on UTC. Change it on the Settings tab.
• Working hours. A new agent is created with them disabled.

Diacritization is applied only to a Fusha voice, whatever the design asks for.

Review the agent after it lands

The design is applied as soon as it passes its checks, so the first time you see it is on the finished agent. Open it, read the Persona tab, and edit anything you disagree with. Every field the build wrote is an ordinary editable field.

Knowledge is attached and the built-in tools are turned on after the agent exists, so check the Knowledge and Tools tabs before you go live and turn anything on that you expected to see.

Limits

Troubleshooting

Microphone access was blocked. Type your answer instead.

The browser refused the microphone. Allow it for this site, or type the answer.

Could not transcribe that. Try again or type it.

Transcription failed on the clip. Record again or type the answer.

We could not hear anything. Try again, closer to the mic.

The transcript came back empty. Record again, closer to the microphone.

Could not start the build. Try again.

Submitting the build failed. This also covers a file that could not be uploaded, since the sources are created before the build is submitted.

Agent design timed out. Try again.

The design did not come back within 15 minutes. A build is never retried for you, because a second design run could produce a second agent. Dismiss the card and start again.

Build failed with a field reason

The design came back but did not pass its checks, and the card shows the exact reason. The ones you can hit are name is required, systemPrompt is required, voiceId … is not in the voice menu, language … disagrees with voice …, speed must be 0.5-2.0 in steps of 0.5, cfg must be 1.0-5.0 in steps of 0.5, numSteps must be one of 16, 32, 48, 64, 128, model … is not in the catalog, llm.temperature is required, llm.temperature must be between 0 and 2 and unknown system tool key …. Dismiss and build again.

Build failed with no reason on the card

The design reported a failure and sent nothing else. Dismiss and build again.

Clearing a failed card

Select Dismiss on the card. Only failed builds can be dismissed; a build that is still running stays on the list until it finishes.

FAQ

Can I close the dialog while it builds?

Yes. That is the intended path. The build carries on without the dialog open and the agents list shows its progress.

What happens if one of my files fails to index?

The build waits up to 15 minutes for your sources, then goes ahead with whatever finished. A file that never indexed is not part of what the model read.

Can I build with no files at all?

Yes. The sources step is optional, and a build with no sources goes straight to the design with no waiting.

Does the agent get the files I uploaded?

Yes. The sources you added during the build are attached to the finished agent as its selected knowledge. Retrieval settings come from the design, so check them on the Knowledge tab before you go live.

Why is the card's name not the name I wanted?

The card name is a placeholder cut from your first answer. The real name is written by the design and appears when the agent is created.

Related

• Create an agent: the template and scratch routes
• What is an agent: the parts the design filled in
• Write the persona: edit the prompt the design wrote
• The agent overview tab: where the finished agent opens