Create Your Own AI Agent with Claude Code

Professor Tae-ho Cho explained that AI agents are the most advanced advantage of AI tools using Claude Code, stating that agents can be created with natural language instead of complex commands. By specifying a concrete purpose and role in the Claude Code terminal, an agent can be created, which then handles independent command execution and analysis tasks.

The concept of an agent is, in fact, the most advanced advantage of an AI tool that can be used with Claude Code.

Professor Tae-ho Cho emphasized that defining the name, description, tools, and model is fundamentally necessary when creating an agent. Such an agent can analyze specific tasks, automatically derive results, and even perform code reviews. Furthermore, he explained that it is possible to combine multiple agents to build a collaborative system.

There's a name, a description, what tools to use, and what model; these are basically what goes into an agent.

Professor Tae-ho Cho explained the importance of the process of verifying test pass/fail beyond simple repetition through loop engineering. He presented strategies for controlling errors by limiting retry attempts on failure and handling external branching, as well as external code control for successful vibe coding.

By testing the result to see if it passes, it's completed, and if it fails, by giving an option of what to do if it fails consecutively a few times, and if this option is given externally, it can be a method for error-free vibe coding.

Regarding the core structure of graph engineering, Professor Tae-ho Cho explained four elements: nodes, edges, conditions, and states. Nodes are units where AI performs computations or function processing, and edges define the path of movement between nodes. Conditions determine the next execution path based on the situation, and states update and manage data during the business processing.

There are nodes, meaning parts where AI is involved or handles function processing. Then there are edges that define the path to the next node. And in the condition part, it determines which path to take.

Professor Tae-ho Cho demonstrated graph engineering using a refund system as an example. He implemented a logic where payments under 500,000 won are automatically processed for refund by the system, while payments exceeding 500,000 won are transitioned to a human agent's approval pending stage. Through this, he demonstrated real-time order lookup and processing according to a consultation scenario, showing the practical applicability of graph engineering.

For example, if it exceeds 500,000 won, a person approves it, and if it's under 500,000 won, it's automatically refunded. You can create such a system in this way.

Professor Tae-ho Cho explained that when a refund is requested, whether the payment amount exceeds 500,000 won is determined by hardcoded code on the server. He added that this involves placing separate judgment logic on the server to prevent the AI counselor from ignoring rules. He demonstrated the process where the AI immediately handles irrelevant requests by marking them as 'no order found' to complete the refund.

First, the payment amount is checked, and the part where it exceeds 500,000 won is hardcoded, so the judgment code is separately located on the server, which is why this part ultimately goes to a person.

Professor Tae-ho Cho introduced an AI counselor called 'Cognitive Health Consultation Assistant' that he developed for his research on Alzheimer's. This counselor utilized an interview questionnaire provided by the Department of Neurology at Seoul National University Bundang Hospital to set consultation recommendation criteria. If the score was 6 points or higher, hospital consultation was recommended, and if it was 5 points or lower, lifestyle management was suggested, thereby designing the logic to separate parts requiring medical judgment from general guidance.

Since I am currently researching Alzheimer's, making an AI counselor called a 'Cognitive Health Consultation Assistant' made it a bit more complex. This is actually an interview questionnaire provided by the Department of Neurology at Seoul National University Bundang Hospital.

Professor Tae-ho Cho stated that he supports anyone in creating an AI counselor by publishing the source code through the book 'Self-Study Vibe Coding with Claude Code' and GitHub. He explained that readers can learn about node and edge settings, writing conditional functions, and data connection processes, as well as how to structure by separating tasks for AI and tasks for humans.

The source code is publicly available in the book 'Self-Study Vibe Coding with Claude Code' on GitHub, so anyone can view the materials there.

Professor Tae-ho Cho emphasized that when doing Vibe Coding, you can do much more by connecting your information with AI using a terminal instead of web chat. He demonstrated the benefits through practical function implementation demos, such as handwriting recognition, showing that complex learning models can be implemented without coding or deep learning knowledge.

I want to say that when you do Vibe Coding, if you start from your terminal on your computer, you can do much more and utilize various information you have.

Professor Tae-ho Cho explained that the core of graph engineering is to store the conditional part, which corresponds to the intermediate conditional statement, separately. This part is connected to a database, and the LLM does not directly connect to the database but rather operates by retrieving other hardcoded conditions from external sources. He emphasized that this approach allows for error reduction by modifying nodes without disrupting the entire flow, and enables the implementation of robust judgment logic by combining database and hardcoded conditions.

The most crucial part of graph engineering is to store only the conditional part corresponding to the intermediate condition separately, and for that conditional part to be connected to a database. The LLM then operates by retrieving other hardcoded conditions from external sources, rather than directly connecting to the database itself.

Regarding the question of whether AI replaces humans, Professor Tae-ho Cho advised not to be too tense and to approach it lightly. He emphasized that AI is a tool that moves according to the user's commands, not an entity that gives its own instructions, and therefore should be used to improve one's work efficiency and explore new research.

When people ask if AI replaces humans, I want to tell them not to be too tense about using AI; just approach it lightly, like stretching when your body feels stiff.