These three methods address different understanding problems. Explanations create a path for following the meaning of a change. Microworlds let people experience states and movements that are usually hard to see. Shared spaces let a team develop understanding together instead of keeping it only in one person's head.
図：説明、マイクロワールド、共有空間という三つの方法が、理解のための技術としてまとめられています。
Figure: Explanations, microworlds, and shared spaces are summarized as techniques for understanding.
This slide shows the speaker reviewing the three methods. What can be confirmed is the three names and the idea of combining them for understanding. The image alone does not prove that each method is effective. Watch the video at 00:17:03.
This does not mean that we should read every line of code one by one. Explanations let us understand a change through its background and purpose. By interacting with a microworld, we can check how states change in a small environment. In a shared space, a team can look at the same document and discuss questions and the reasons behind decisions. In other words, there are ways to understand a system besides reading its lines in order.
Understanding here does not mean memorizing every detail. It means being able to explain the important parts of the work we are involved in and the relationships between them. It also matters that we can connect that understanding to our next decision or idea.
AI agents make it faster to write code and move changes forward. But speed is not the same as understanding. The more we delegate to an agent, the more our work may look finished while we lose opportunities to grasp what happened.
If we leave this state alone, it can lead to the cognitive debt discussed earlier. At first, postponing understanding may seem harmless. But as changes accumulate, more parts become impossible for us to explain ourselves. As a result, it becomes harder to make decisions in the next work loop or generate new ideas.
Understanding is also necessary for creative participation. When we understand one loop, we can carry that understanding into the next loop. We can notice new problems there and combine several ideas. In contrast, simply looking at the result and approving it with “It seems fine” may not develop enough understanding to create the next idea.
Suppose an agent makes a change and all the tests pass. If a person only approves the result, they may not know which parts the change affects. If they read an explanation and, when needed, check the behavior in a small environment, they can more easily propose the next change themselves. This example is a supplement to make the speaker's idea easier to understand.
Understanding is not only an individual problem. If team members use shared names for system parts, screens, and ideas, they can talk about the same things. When questions and interactions with agents are shared, the team can also see problems and reasons for decisions that one person could not see alone.
This is different from two people each consulting a separate agent and bringing back only the results. In a shared space where the conversation is visible, one person's question can create another person's question. If everyone can view and comment on a plan in the same document, they can discuss it in context instead of passing summaries back and forth. Through this process, the team builds a shared model—a common way of seeing how the system works.
The speaker uses the word “defend” because we should not treat understanding as an automatic result of work. In an AI-assisted environment, speed and delegation—letting other people or agents handle work—are highly valued. Therefore, if we try to gain understanding later, the opportunity may not still be there.
That is why people need to create deliberate entry points to understanding. We can generate an explanation of an agent-made change. When needed, we can try the change in a small, interactive environment. We can add comments to plans used by the team and share questions and decisions. This lets us delegate work to AI without completely removing people from the work loop.
Here, “defend” does not mean doing every task only by hand. It means deliberately combining the parts we delegate so that we can move faster with the parts where people understand the system and make judgments and create ideas. The speaker's conclusion is a call to build ways of preserving understanding into AI-assisted work.