AI Agents Must Be Managed Like People

A phenomenon has emerged where AI agents introduced into an organization lose functionality overnight due to the departure of the responsible employee, reverting to previous work methods (such as Excel). Kim Deok-jung, Director of Pub AI Research Institute, pointed out that an agent meticulously created over a year became dysfunctional in less than a day when the person in charge left, demonstrating the vulnerable sustainability of AI agents. This phenomenon indicates potential problems that can arise if a continuous management and operating system is absent when AI agents are introduced.

It has been argued that managing AI agents with a lifecycle similar to human employees is necessary to prevent failures. Director Kim Deok-jung emphasized that just as human employees go through recruitment, training, performance evaluation, and retirement, AI agents also need a full process of introduction, training, evaluation, and dismissal. He particularly stated that agents consuming only resources (tokens) without performance should be decisively dismissed, highlighting the importance of establishing strict criteria for effective AI operation.

Director Kim Deok-jung explained that many attempts are being made to integrate AI into processes, mentioning a paradigm shift where AI is becoming central to work.

Director Kim Deok-jung explained that the term 'agent' is the first instance of technology being personified, literally meaning a proxy or someone who performs tasks on another's behalf. He emphasized that an agent is not a tool that learns and is used by a user, but rather independent technology that acts on behalf of the user, suggesting the increasingly sophisticated role of AI agents.

He explained that by clearly identifying the organization's 'pain points' and deploying agents optimized for them, work efficiency can be maximized.

AI agents can be broadly classified into four types based on organizational needs. First, chatbot agents handle user inquiries, such as HR or purchasing regulations. Second, copilot agents assist with productivity tasks like document creation. Third, workflow automation agents automate work processes through intelligent task classification and connection (RPA). Finally, independent agents are the most advanced form, autonomously executing tasks when specific conditions are met.

First, the agent must clearly understand the task objectives and success criteria. Second, it must be able to utilize various models and inference methods. Only by meeting these technical and operational requirements can agents achieve the expected performance.

Director Kim Deok-jung stated that agents must be good at memorization, learning from user conversations to improve and generate outputs without repeating errors. This includes knowledge, memory, and context understanding. Additionally, the ability to integrate with external systems using MCP (Model-Controller-Program) technology to utilize skills and tools is essential, citing an example of HR management system integration.

Human control and management systems are essential for AI agent operation. Director Kim Deok-jung emphasized the importance of execution authority and approval functions as the fifth essential condition. Through the 'Human-in-the-loop' concept, agents can autonomously determine areas requiring human approval, thereby ensuring the accuracy of results and enabling auditing. Furthermore, appointing an operations manager and establishing continuous improvement cycles are critical.

Four key indicators were presented for evaluating AI agent performance. Director Kim Deok-jung stated that when selecting an agent for individual or common organizational tasks, four criteria must be checked: speed, scale, quality, and learnability. Speed refers to reducing task processing time, and scale refers to scalability across the entire organization. Quality involves verifying the accuracy of results and compliance with plans, while learnability indicates the agent's continuous improvement capability. These indicators allow for objective measurement of the agent's impact.

Director Kim Deok-jung emphasized that an agent can only be created by commanding it to perform specific unit tasks, such as writing documents, classifying requirements, or generating something. This means that efficiency can be increased by concretely dividing tasks and developing agents suitable for each task.

Seven key conditions were presented to clearly define tasks for AI agents. First, 'Trigger' defines when a task begins, and 'Input Condition' specifies the information needed to perform the task. Second, 'Decision Rules' set the agent's judgment criteria, and 'Tool Usage Capability' determines whether necessary tools can be integrated. Third, 'Output Result' defines the deliverables generated upon task completion, and 'Responsible Person' designates the individual ultimately accountable for the task. Finally, 'Exception Condition' specifies how the agent should respond when situations deviate from the norm. Only when these seven conditions can be specified can a task be considered a single unit for AI agent application.

Various types and platform options exist for building AI agents. Director Kim Deok-jung mentioned that general-purpose platforms like Claude Cowork or Codex can be used directly. He also explained that leveraging low-code build methods, such as Microsoft's Copilot Studio, or using agent platforms provided by cloud platforms is possible. Lastly, he added that companies can build their own agents using open source, emphasizing the importance of choosing based on each organization's situation and needs.

Director Kim Deok-jung emphasized the importance of enterprise-specific features such as authorization management, account management, and approval management, indicating the complexity of operating enterprise AI agents where security is essential.

When implementing AI agents enterprise-wide, a systematic management strategy is needed. Director Kim Deok-jung stressed that if agents are used by all employees and impact customer service, a proper cloud platform or internally developed platform should be used. He also explained the importance of clearly distinguishing agents into personal, organizational, and enterprise levels through a central organizational body within the company and defining the requirements for each level. This implies that agent implementation should be approached from a portfolio management perspective.

Indiscriminately proliferating AI agents can lead to financial liabilities for a company. Director Kim Deok-jung warned that agents are 'money-eating hippos' and that in the not-too-distant future, token and infrastructure usage fees comparable to labor costs will have to be paid. Therefore, he emphasized that continuous proliferation of agents like amoebas becomes a liability from a corporate perspective, and a compact and efficient operation is the ideal model.

Clear principles for separation and consolidation are necessary for efficient AI agent management. It is reasonable to separate agents if their success goals, access data, system security policies, approvers, or authority fundamentally differ. Conversely, Director Kim Deok-jung explained that agents with low user demand can be consolidated to increase efficiency. Establishing such clear guidelines is crucial to prevent indiscriminate agent proliferation and avoid them becoming liabilities.

The era of 'agent armies' is dawning, with agents performing various roles beyond simple task execution. Director Kim Deok-jung explained that task-centric agents, like a 'meeting agent' that automatically generates meeting minutes from recorded audio, are fundamental. Furthermore, he emphasized that agents for monitoring proper agent operation, managing agent performance, and identifying improvements are needed and being applied. This signifies the growing importance of a seamless collaborative system among diverse AI agents.