Agent-style apps

Agent-style apps

This guide introduces how to build an app using agents. Agents can execute tasks that an LLM cannot directly answer/perform by using simple instructions and external integrations. Integration with MCP (Model Context Protocol) servers and tools is supported, allowing use of various MCP ecosystem tools as well as internally added document management tools. Integrate multiple tools to automate workflows. The agent will use the necessary tools according to the request.

Getting started: Building an agent app

Enter the app management page, click the "Create new app" button, and then select the "Create agent" card. You will be taken to the agent builder page as shown below.

1. App information: name, icon, description

• Like existing answer-type and conversational apps, you can set the app's name, description, and icon. Your settings will be displayed in the admin dashboard and Works.

1. LLM model settings and instructions

• LLM model: Select the desired model.

• Instruction: Enter instructions (system prompt) for the Agent. Assign persona and role, and provide specific guidelines to follow.

• Variables: Specify variables like {variable_name} to substitute values. Using the "Add variable" button, you can require users to input values for variables when the app runs, or use system variables. Variables set here are the same as variables in conversational apps. For more details on using existing variables, here.

General variables (when running the app, the configured value is assigned to the variable; if no value is set, an empty value is assigned.)

1. Tools

• You can integrate MCP servers and add tools (MCP Tools). From simple tools like time checks and web searches to professional automations via integration with various third-party services' MCP Servers.

• Click the "Add" button to view MCP servers and tools integrated per project as shown below. Select the tools you need and add them to the agent. *If you cannot find the required tool in the product, please contact your representative.

• After adding, click the integration/configure button for each server and tool as shown below to enter the information or detailed argument values required for integration. Tool arguments are generally automatically filled by the agent based on the user's query and the identified context. Cases where the user enters values directly are considered exceptions, and those values are filled as overrides that always execute.

Enter the API Key to integrate the MCP server. Authentication methods vary by server, such as OAuth.

You can view the arguments to use the tools (MCP tools).

1. Conversation scenarios

• For smooth app use, you can specify the conversation start message, the message shown when a model call fails, and the variable value request message (used only when "Add user input variable" is selected in variable #2 above).

1. Start preview

• Click the button to run the app in the preview area on the right. It becomes active once all required fields are filled. If changes occur afterward, the button changes to "Update preview", allowing you to run the preview again based on the updated content.

Example app

Based on the above, a very simple "return time by region" app would look like the following.

A look at the built app

View of the executed preview

Publish - advanced users

If you want to create a more complex and specific-flow conversational app using the agent you built, use it in the following way.

• Click the More (three dots) button, then select "Publish agent only." In this case, the app will not be published andyou can use the built agent in the conversational app's LLM execution nodeonly (Even if you choose "Publish app", you can still use the agent in the LLM execution node. However, the above "Publish agent only" does not publish the app on Works.) This allows you to create more customized messages and flows as shown below.

Version control

In the upper right, view the history of saved or "published" versions and edit or 'publish' each version.

• Preview is supported for each version.

• You can create/edit a new version while leaving the published version as is.

• Select a specific past version or the currently editing "Current", then choose "Publish" to publish without changing the existing app to a 'draft' state.

Change to drafting

You can change the app to draft status from the app management page. The changed app will no longer be exposed on Works and will not appear in the agent list of the LLM execution node.

Change to drafting - advanced users

Even after entering the app builder page, selecting the currently published version allows you to change the app and agent status to "draft."

• If the app is "draft," it will not be visible or accessible on Works, and if the agent is "draft," it will not be exposed in the LLM execution node and cannot be used.

• Because other versions may be eligible for 'publishing,' the "Change to draft" button is shown only when entering the currently published version.

If this agent is being used in another conversational app, the agent cannot be changed to draft. However, as shown below, the app'only' can be changed to draft.

• As above, this option is displayed only when entering the currently published version.