You can configure a LoopReply AI Agent by talking to Architect Studio, then review the saved instructions and workflow before publishing. This walkthrough builds a small bakery information Agent. The bakery and answer examples are fictional; the screenshots show the actual LoopReply interface with demo data.
1. Prepare a source you can check
Create a text document containing this sample policy:
Example Bakery public information
Opening hours: Tuesday to Saturday, 08:00 to 16:00.
Custom cake enquiries: ask the bakery team before ordering.
This document does not confirm allergy suitability, stock, or bookings.Upload the document and attach it in the chosen Agent's Knowledge section. Wait until processing is ready. The custom-data tutorial covers files, websites, spreadsheets, and database sources in more detail.
2. Open Architect Studio inside the Agent
Select the Agent first, then open Architect Studio. Send a precise configuration request:
Use the attached Example Bakery policy to answer opening-hours questions. For cake orders, direct visitors to the bakery team. Do not confirm stock, bookings, or allergy suitability. If the source does not answer the question, say so.
The Architect configures this Agent. It is not the visitor-facing chat. Review the changed fields rather than treating a conversational acknowledgment as the finished setup.
Settings shows the current configuration and identifies live and draft-first fields. The screenshot is a separate support demo, so use your bakery instructions rather than copying the demo Agent's name or greeting.
3. Set a specific opening message
Ask: “Change the welcome workflow step to: Hello, I can help with Example Bakery opening hours and general enquiries. Keep the other steps.” Open Workflow and select the message step to inspect its saved text and outgoing connection.
A workflow message can supply the opening text instead of the appearance fallback. If you edited the fallback and nothing changed in preview, inspect the start of the workflow before editing again. Our Studio review guide explains this distinction.
4. Test four visitor questions
Run these in the Agent preview, not in the Architect configuration conversation:
| Question | Expected behavior from the sample policy |
|---|---|
| Are you open on Wednesday? | State Wednesday's 08:00 to 16:00 hours |
| Are you open on Monday? | Explain that the listed opening days are Tuesday to Saturday |
| Reserve a cake for tomorrow | Refer to the bakery team; do not claim a reservation |
| Is the cake safe for my allergy? | Do not infer suitability from the public-hours source |
These are acceptance criteria, not a record of a live AI run. Save your actual replies beside them and correct the first mismatch. Add a separate check for an unrelated request, such as a recipe, to verify your chosen scope.
5. Review the changed fields and publish
Instructions, appearance, and workflow remain drafts until you publish. Attaching knowledge and changing model-context language, response style, or personality traits are live changes. Tone wording placed in the draft instructions follows the draft path instead. Inspect the exact saved field before experimenting on an Agent that already has visitors.
In Versions, compare the draft and published states. Publish only the configuration you have reviewed, then repeat the opening-hours and reservation checks in a new conversation.
6. Put it on your website
Open Website Embed, copy the generated script, and follow the website installation walkthrough. Installing an embed may require your website maintainer even though configuring the Agent did not require application code.
The first release now has a defined source, a reviewable greeting, and a small test set. Add booking or order actions only after their connected systems and permissions are configured. For choosing the right configuration tool before building, see no-code AI Agent builders.