[Edaily Reporter Kim Hyun-ah ] Kakao(035720)has established a system where multiple AI agents collaborate through the company’s internal AI assistant, “Workmate,” to handle tasks ranging from coordinating meeting schedules to reserving conference rooms.

A key differentiator is that the AI saves an action plan and obtains approval before making changes to actual business systems; once approved, the AI executes the confirmed plan without reevaluating the details. The system is designed to automate tasks while preventing incorrect execution and the misuse of permissions.

At the Kakao developer event “if(kakao)2026,” held on the 14th, developers Kim Dong-hoon (JackJack) and Yoon Ho-se (Heyo) introduced the design and operation of the in-house AI assistant “Workmate:.”

When an employee requests a meeting, “WorkMate:” verifies the attendees and reviews their work schedules and existing calendars to find a mutually available time. It then checks meeting room availability and issues a confirmation card containing the proposed plan. Once the user approves it, the actual reservation is processed.

Developer Kim Dong-hoon explained, “Even to schedule a single meeting, you have to find people on the organizational chart, check their attendance status to see if they’re in the office or working remotely, and match available times on their calendars before you can reserve a room through the meeting room system.” He added, “WorkMate navigates these systems on the user’s behalf.”

AI Roles Divided by Task… Connected via A2A

Rather than having a single AI handle all tasks, “WorkMate:” divides AI

responsibilities by

work

area

, such as scheduling and meeting room reservations. An AI that coordinates the entire workflow breaks down requests into specific tasks and assigns them sequentially to the respective AI agents.

Initially, the AIs were organized by existing systems, but considering the need to check both attendees’ work status and calendars simultaneously, attendance and scheduling were combined into a single task area.

Developer Kim stated, “The key is that we redesigned the boundaries based on the tasks and rules that require joint decision-making.”

The system uses the A2A (Agent-to-Agent) standard for AI-to-AI task handoffs. Once “agent cards”—which contain each AI’s responsibilities and invocation methods—are registered, the coordinating AI reviews them to select the appropriate AI. This allows AIs developed by different departments to be interconnected according to a common standard.

A feature has also been implemented to link the titles employees typically use to their internal accounts. WorkMate infers titles by analyzing employee profiles, self-introductions, and work-related conversations, then searches various information—such as account IDs, real names, and pronunciations—to find the user matching the request.

Development Tools for Claude Code Also Provided

Kakao provides skills and development documentation for “Claude Code”—Anthropic’s AI development tool—so that other departments can develop business-use AIs and connect them to “WorkMate:”. Developers can use a common structure to create a basic framework and add necessary business functions. It also supports an execution library that can be used in a Java Spring environment.

However, developed AI systems cannot be connected immediately. Kakao uses a validator to verify authentication information, request and response formats, and approval status. For AI systems with write capabilities, it also checks whether duplicate executions are prevented through approval procedures and request retransmissions.

Processing via Approved Execution Plans… Permission Restrictions by AI

“WorkMate:” does not immediately execute tasks that modify information in actual business systems. It saves scheduled execution plans to a confirmation card, and once the user approves them, it resumes the task using predefined tools and parameters.

Developer Kim explained, “It pauses rather than immediately calling the scheduled API (Application Programming Interface).” This is intended to minimize discrepancies between what the user has confirmed and the actual execution by ensuring the AI does not re-evaluate the scheduled details after approval.

If conditions change—such as the addition of a participant—the existing task is halted, and a new confirmation card is issued with the updated details. The modified reservation must be reapproved.

Permissions are also restricted on a per-AI basis. Each AI is provided with a dedicated token containing only the permissions necessary for the task it is called to perform. For example, if the scheduling AI attempts to call the reservation function without the necessary permissions for booking a conference room, the system rejects the request.

Developer Kim emphasized, “Only the permissions necessary for the target task should be granted.” The goal is not merely to divide the roles of AI systems by task but to apply approval procedures and permission controls together to prevent incorrect execution, abuse of permissions, and duplicate processing.