Table of Contents
Artificial Intelligence Translates the Language of Animals
Kim Kyung-jin, Attorney at Law
The Story of AI Learning to Listen to Whales, Dolphins, Birds, and Bees
Twelve chapters on how AI listens to dolphins, sperm whales, humpback whales, birds, and bees to find rules in their sounds, and what this technology means for its risks and for animal rights. Written in simple sentences a child can read, with verified sources in every section.
Table of Contents
AI Deciphers Ancient Scripts
Kim Kyung-jin, Attorney at Law
Ancient Records Revived by AI
In twelve chapters, this book explains how AI revives records once unreadable, from burned scrolls and wooden slips buried in mud to broken clay tablets. It covers virtual unrolling at Herculaneum, virtual collation of oracle-bone texts, reading Silla wooden tablets, computational analysis of undeciphered scripts, and multispectral archives, with verified references for each chapter.
Table of Contents
Artificial Intelligence for New Materials Design and Rocket Propulsion Engineering
Kim Kyung-jin, Attorney at Law
AI Potentials, Self-Driving Laboratories, and Physics-Informed Machine Learning (PIML)
Ten chapters on how artificial intelligence is changing new materials and rocket propulsion: atomic simulation, generative models, self-driving labs, high-temperature alloys, metal 3D printing, combustion, cooling design, and engine diagnosis and control. Written without equations, with verified sources in every chapter.
AI Library
The Age of Autonomous Scientific Discovery
Kim Kyung-jin, Attorney at Law
AI Scientists and Self-Driving Labs
This book follows how AI scientists and self-driving labs are changing the way science generates and verifies claims. It covers literature-based discovery, natural-language protocols translated into robot commands, multi-agent research systems, closed-loop laboratories, materials search, the verification gap, chains of evidence, research harnesses, journal ethics, and legal responsibility.
AI Library
A New Era of Life Sciences Opened by Artificial Intelligence
Structural Proteomics, Genomic Foundation Models, Autonomous Laboratories, and Global Governance
Kim Kyung-jin, Attorney at Law
This book is a research volume compiled with artificial intelligence. A human selected the materials and structured the work, while AI models drafted the sentences and cross-checked the facts.
AI Library
The Double Structure of Digital Sovereignty
Europe’s Departure from Palantir and the Chains of American Big Tech
Kim Kyung-jin, Attorney at Law
This is a record of 2026, when European intelligence agencies and defense ministries began removing analytics tools from America’s Palantir. It covers the replacement decisions made by France’s General Directorate for Internal Security (DGSI), Germany’s Federal Office for the Protection of the Constitution (BfV), and the Netherlands Ministry of Defense; the incident in which US export controls severed an ally’s ac…
New English Edition
Artificial Intelligence in Horticulture
Kim Kyung-jin, Attorney at Law
Across five chapters and ten sections, this book examines computer vision for crop diagnosis, harvesting robots and autonomous field systems, smart greenhouses and digital twins, precision irrigation and supply-chain quality control, high-throughput phenotyping, and predictive breeding.
New English Edition
Artificial Intelligence in Food Crop Agriculture
Kim Kyung-jin, Attorney at Law
Across six chapters and eighteen sections, the book examines digital agricultural infrastructure, remote sensing, crop diagnosis, yield forecasting, precision irrigation, genomics, molecular breeding, agricultural robotics, climate-smart agriculture, and global food security.
New English Edition
The Future of Forestry and Agroforestry
Kim Kyung-jin, Attorney at Law
Driven by Artificial Intelligence and Digital Innovation
Across five chapters and fifteen sections, the book follows satellites, drones, LiDAR, digital twins, forest-specific language models, wildfire and pest forecasting, forestry robotics, agroforestry, timber traceability, and forest carbon markets.
New English Edition
Smart Livestock Farming: AI Enters the Barn
Kim Kyung-jin, Attorney at Law
Sensors listen, cameras watch, and artificial intelligence helps farmers decide.
Across five chapters and fifteen sections, the book follows precision livestock farming from animal health and reproduction to robotic milking, virtual fencing, digital twins, methane reduction, welfare, and data ownership.
Table of Contents
Han Dong-hoon, Busan Buk-gu Gap: A Record of the 100 Days Before and After the Election (Mar. 26-Jul. 3, 2026)
Kim Kyung-jin
Table of Contents and 13 sections
From March 26 to July 3, 2026, this record follows the spring after expulsion, the Busan Buk-gu Gap by-election, victory as an independent, and the first bill submitted in the National Assembly.
Table of Contents
Artificial Intelligence and Medicine
Kim Kyung-jin, Attorney at Law
AI in clinical care, hospitals, education, and research
AI in medical imaging, risk prediction, treatment planning, hospital operations, education, and research, with patient safety, privacy, and accountability.
[AI Library] Chapter 24: Sub-agents: The Art of Delegating Work
Mastering Claude Code
Chapter 24: Sub-agents: The Art of Delegating Work
Kim Kyung-jin
Mastering Claude Code
The Relationship Between Main Session and Sub-agents
I opened Claude Code in a terminal window. When I invoked the Carousel Skill, a small agent indicator appeared on one side of the screen. "I will plan carousel slides. I am using an agent." Instead of the main session composing the slides directly, I delegated the task to a separate sub-agent (Subagent).
While the sub-agent worked diligently to structure the slides, the main session could continue with other tasks.
This scene compresses the essence of what a sub-agent is.
When you open a new session in Claude Code, the one you are speaking with is the main session. The main session has its own context window size, builds up a conversation history, and oversees the overall project. But what if this main session had to perform every task directly?
If you process code review, research, test writing, debugging, and documentation all within a single context, the window quickly becomes saturated. Context rot begins, and the context of previous conversations becomes murky.
A sub-agent is a structure built to solve this problem. If the main session is a project leader, then a sub-agent is a specialist team member. The project leader manages the big picture, and the specialist handles the concrete work.
[Figure 24-1] Relationship diagram between main session and sub-agents: The main session (project leader) delegates work to six sub-agents,code reviewer, builder, debugger, test runner, architect, and researcher.
Let me look more closely at how sub-agents work.
A sub-agent sleeps until awakened. Most of the time it is as if it does not exist. The moment the main session calls it,"Wake up, do this task for me",the sub-agent awakens with a completely new context. It can use its own model, has its own set of tools, and focuses on its own purpose.
The key point is that it has an independent context. It does not share the conversation history of the main session. It receives only the prompt the main session sent, performs the task independently, and returns just the result.
The ability to use a different model is also important. Even if the main session is running on Opus, a sub-agent can work with Haiku or Sonnet. This opens the door to fine-tuning both cost and speed.
And parallel execution is possible. In one demonstration, when I requested "Do research on SME AI adoption and enterprise AI adoption simultaneously," the main session dispatched two research sub-agents at the same time. Each agent completed its investigation independently and sent the results back to the main session. The research returned included an SME AI adoption brief, an enterprise AI adoption brief, and even the intersection of both.
If you compare this structure to an organization, the main session focuses on making plans, distributing work, and consolidating results. The actual execution of specialized tasks falls to the sub-agents. As a result, the main session's context stays clean.
The Custom Bakery Analogy
Imagine you are preparing for a party. You really need delicious cupcakes.
There are two options. You can go to a large supermarket and grab mass-produced cupcakes. It is fast, cheap, and convenient. But the taste is just ordinary. Alternatively, you can visit a specialty bakery in town and place a custom order. "Put vanilla cream with a touch of lavender, and place edible flowers on top." It takes longer, but the quality of the result is incomparable.
[Figure 24-2] Custom Bakery Analogy: Supermarket cupcakes (general processing) vs. custom bakery (specialized sub-agents) comparison.
A sub-agent is a custom bakery.
The main session handling all tasks directly is similar to buying cupcakes from a supermarket. Processing multiple tasks with a general set of capabilities is faster, but there are limits to the quality of each task. When code review, research, test writing, and debugging are all mixed within a single context window, it becomes difficult to focus deeply on any one.
It is different when you delegate to a sub-agent. A "code review specialist" sub-agent has a system prompt designed only for code review, a tool set optimized for code review, and a context focused on code review. Just as a custom bakery pours all its craftsmanship into a single cupcake, a sub-agent concentrates all its capability on the delegated task.
There is one more point to note in this analogy. When you place an order at a custom bakery, you communicate clearly what you want. "Vanilla cream, lavender scent, edible flowers." The same goes when delegating a task to a sub-agent. The prompt the main session sends to the sub-agent is the order form. The clearer the order form, the higher the quality of the result returned.
And a custom bakery can take repeat orders. "Those cupcakes you made last time,I would like 20 more this time." A sub-agent works the same way. Once you build one well, you can reuse it across projects and throughout the entire organization.
Five Reasons to Use Sub-agents
Now that we understand what a sub-agent is and how it operates, let us look at the specific situations in which a sub-agent shines.
It is to preserve context. The most precious resource to manage when using Claude Code is the context window. Context rot is a real threat. The operator's role is to manage the context window as efficiently as possible, and sub-agents are one of the key means to do so.
Process large volumes of data, perform broad research, and distill from those results only the core information the main session actually needs, sending that back. In one demonstration, a sub-agent performed work equivalent to approximately 40,000 tokens of text, yet the main session remained at 29,000 tokens consumed.
The content of the 40,000 tokens the sub-agent processed did not pollute the main session's context.
It is to apply constraints. The main session should be able to call a wide range of tools. However, things are different when performing risky tasks,GitHub Actions or certain code base operations. By assigning only a limited set of tools to a sub-agent dedicated to such work, you can prevent unintended damage.
By separating the main session's powerful authority from the sub-agent's limited authority, this is a strategy to increase the safety of the entire workflow.
It is to reuse configurations. A sub-agent is defined in a single markdown file. If you copy that file to another project, you can use the identical sub-agent immediately. Thinking at an organizational level, if you build one sub-agent that excels at project management or quarterly planning, the entire organization can benefit from it.
The principle is similar to skills, but sub-agents add the dimensions of independent context and model selection.
It is for specialization. AI produces better results the more specific and narrow its purpose. Multiple small agents each dedicated to a single stage are better than one multi-purpose agent that does planning, building, testing, documentation, research, and review all at once. The main session just needs to synthesize those results. The principle "Tiny agents win" has been proven repeatedly in practice.
It is to reduce costs. Since you can assign a different model to each sub-agent, you do not need to use Opus for every task. If you assign Haiku to a sub-agent that needs quick research and Sonnet to one that needs code generation, you can save both time and token costs.
[Table 23-1] Summary of reasons to use sub-agents.
Judgment Criteria for When to Delegate and When to Do It Yourself
Once you know the advantages of sub-agents, the temptation arises to delegate every task. But doing so falls into over-engineering. You need criteria to distinguish when to delegate and when to handle it directly.
Tasks the main session should handle directly:
Tasks that require back-and-forth conversation. When you request "Fix this part," see the result, then ask "No, go this direction instead," and repeat that iterative interaction, it is more efficient to do it directly in the main session. If you delegate to a sub-agent, you must start from scratch in a new context each time, which is actually less efficient.
Tasks requiring quick fixes are the same. Changing a single variable name, fixing a typo, or modifying simple logic,calling a sub-agent for such work makes the overhead of awakening the sub-agent and receiving the result larger than the task itself.
Tasks that require shared context,work that must be judged based on the conversation history so far,should also be handled by the main session. This is because a sub-agent does not inherit the conversation history.
Even in situations where low latency is required, the main session has advantages. Calling a sub-agent, performing the task, and receiving results inevitably consume time.
When you should delegate to a sub-agent:
Self-contained work. Tasks like "Analyze this codebase and find security vulnerabilities" are good candidates for delegation, because the sub-agent receives input, performs the work independently, and returns results.
Sub-agents are also appropriate for work requiring tool restrictions. You can grant a specific sub-agent read-only access, or limit it to use only certain MCP connection servers.
When you need only summarized results, that signals delegation. Even if a sub-agent processes vast data, the main session receives only the key summary. This is the core mechanism for preserving context.
There is also the choice between foreground and background execution. Running a sub-agent in the foreground means the main session waits for it to complete, while running it in the background allows the main session to continue with other work. This choice depends on the task's urgency and dependencies.
[Figure 24-3] Delegation decision flowchart: "Do you need back-and-forth conversation?" → Yes: Main session / No → "Is it self-contained?" → Yes: Sub-agent / No → Main session]
One critical distinction must be made clear. Sub-agents and agent teams have similar names and appear to share concepts, but they are fundamentally different. A sub-agent is a one-way relationship. The main session sends input, the sub-agent performs the work, and returns results to the main session. Sub-agents cannot communicate with each other.
Even when they run in parallel, each operates independently.
Agent teams are bidirectional. Members have a shared task list, can send messages to each other, and assign work to one another. That is collaboration on a completely different level and will be covered in detail in the following chapters.
So far, we have built the conceptual foundation of what sub-agents are, why they are needed, and when to use them. Now it is time to examine how to implement this concept in practice.
Kim Kyoung-jin, Attorney and AI Specialist
AI legal policy specialist; former member of the National Assembly; author of multiple works
If this book has been part of your life, even briefly, please support the next story coming into the world.
(Voluntary support requested. Account: NH Bank 302-1096-0948-81 Account holder: Gyeong-jin Kim)
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.