At the if kakao developer conference held on the 13th at Kakao AI Campus in Yongin, Gyeonggi Province, Kakao unveiled model, service and operational plans for adding personalized AI agents to KakaoTalk. The company said a service used around the clock by roughly 50 million monthly users must address speed, cost, reliability, privacy and control over execution permissions.

Kakao manages cost and response speed by assigning models of different sizes according to request difficulty. Small language models handle repetitive, clear requests; on-device models handle tasks that can run on a smartphone; and frontier models handle difficult requests. Flagship models handle most server-based requests. An “AI conductor” interprets requests, selects suitable models and tools, and sets the execution order. Kakao also uses MOPD training to transfer the capabilities of models strong in specific fields to smaller models.

Noh Byeong-seok, Kakao’s unified foundation model performance lead, said Kakao’s Korean-focused tokenizer made sentence processing 1.6x more efficient than global competitors. Response generation became 3.9x faster on average, while model compression techniques cut memory use by about 70%. Kakao said the amount handled on the same GPU increased about 1.6x. The company is developing services based on Kanana-o that understand text, images and speech together and respond by voice, along with speech technology that conveys emotion and intonation. “Kanana Live” is intended to listen and respond before the other person finishes speaking, and to interpret spoken Korean into English.

Agentic AI is designed to complete requests through real actions. For a restaurant recommendation, for example, an agent could check available reservation times, then book and share the reservation after the user approves. Kakao said evaluation is shifting from answer quality to whether the requested task was actually completed. That requires identifying intent, choosing tools, checking results across multiple steps and revising plans. The AI conductor breaks requests into smaller tasks and coordinates specialist agents for search, recognition and understanding.

Song Jae-ha, Kakao’s chief technology officer, said personalized services need memory that carries a user’s information forward and a platform that acts within permitted limits. Project Mosaic is technology intended to carry user memories and conversation context across KakaoTalk and multiple agents. Kakao records verified facts from user behavior separately from AI interpretations, along with their sources and dates. A single purchase of children’s shoes, for instance, does not establish a preference because the item may have been a gift for a niece or nephew. Purchase history may support reordering, while delivery information may become invalid once delivery is complete, so the system also considers how long information remains valid.

Before sharing memory with another service, the system checks the scope of consent. Information approved for use only in one service is not passed automatically to others; the system determines whether the requesting service may use it and shares only permitted information. Before acting, it checks the data’s intended use and whether consent was given. The organization responsible for the data reviews incorrect information. The system must also block disallowed actions, such as following an instruction to switch to a different product when the user asked to reorder one bought previously. Kakao said it is building a system to verify not only final answers but also task plans and execution, stopping and asking the user when an action exceeds the permitted scope.

Kakao also cited service reliability as a challenge. Song said situations in which traffic reaches 7x to as much as 18x normal levels recur across Kakao services. Kakao is preparing to allocate server resources according to demand and connect external cloud servers to its own systems so resources can be added when traffic surges. Song said what matters is whether an order is confirmed under the conditions the user set, rather than whether the AI produced a good answer.

At the if kakao developer conference held on the 13th at Kakao AI Campus in Yongin, Gyeonggi Province, Kakao unveiled model, service and operational plans for adding personalized AI agents to KakaoTalk. The company said a service used around the clock by roughly 50 million monthly users must address speed, cost, reliability, privacy and control over execution permissions.

Kakao manages cost and response speed by assigning models of different sizes according to request difficulty. Small language models handle repetitive, clear requests; on-device models handle tasks that can run on a smartphone; and frontier models handle difficult requests. Flagship models handle most server-based requests. An “AI conductor” interprets requests, selects suitable models and tools, and sets the execution order. Kakao also uses MOPD training to transfer the capabilities of models strong in specific fields to smaller models.

Noh Byeong-seok, Kakao’s unified foundation model performance lead, said Kakao’s Korean-focused tokenizer made sentence processing 1.6x more efficient than global competitors. Response generation became 3.9x faster on average, while model compression techniques cut memory use by about 70%. Kakao said the amount handled on the same GPU increased about 1.6x. The company is developing services based on Kanana-o that understand text, images and speech together and respond by voice, along with speech technology that conveys emotion and intonation. “Kanana Live” is intended to listen and respond before the other person finishes speaking, and to interpret spoken Korean into English.

Agentic AI is designed to complete requests through real actions. For a restaurant recommendation, for example, an agent could check available reservation times, then book and share the reservation after the user approves. Kakao said evaluation is shifting from answer quality to whether the requested task was actually completed. That requires identifying intent, choosing tools, checking results across multiple steps and revising plans. The AI conductor breaks requests into smaller tasks and coordinates specialist agents for search, recognition and understanding.

Song Jae-ha, Kakao’s chief technology officer, said personalized services need memory that carries a user’s information forward and a platform that acts within permitted limits. Project Mosaic is technology intended to carry user memories and conversation context across KakaoTalk and multiple agents. Kakao records verified facts from user behavior separately from AI interpretations, along with their sources and dates. A single purchase of children’s shoes, for instance, does not establish a preference because the item may have been a gift for a niece or nephew. Purchase history may support reordering, while delivery information may become invalid once delivery is complete, so the system also considers how long information remains valid.

Before sharing memory with another service, the system checks the scope of consent. Information approved for use only in one service is not passed automatically to others; the system determines whether the requesting service may use it and shares only permitted information. Before acting, it checks the data’s intended use and whether consent was given. The organization responsible for the data reviews incorrect information. The system must also block disallowed actions, such as following an instruction to switch to a different product when the user asked to reorder one bought previously. Kakao said it is building a system to verify not only final answers but also task plans and execution, stopping and asking the user when an action exceeds the permitted scope.

Kakao also cited service reliability as a challenge. Song said situations in which traffic reaches 7x to as much as 18x normal levels recur across Kakao services. Kakao is preparing to allocate server resources according to demand and connect external cloud servers to its own systems so resources can be added when traffic surges. Song said what matters is whether an order is confirmed under the conditions the user set, rather than whether the AI produced a good answer.