If you only learn how to ask well, you get one-sentence answers. The real work — the kind that pays — is decided by your ability to set the whole context. We'll walk through that shift today.
Most people who've used AI for over half a year run into the same wall — their prompts have clearly gotten better, but the moment the work gets bigger, AI starts tripping in strange ways. Why? The answer is simpler than you think. You trained yourself on asking well, but never on setting the table.
This essay walks through that shift from start to finish. If you've just started using AI, you can still follow along — we'll go slowly. Today's example is Claude's sub-agent feature, but the principle applies to any AI tool, in any era. Two or three years from now, even when "sub-agent" is renamed, the spine of this essay still holds.
Let's start with the principle. Work has a question and an environment. The question is "please do this." The environment is "everything you need to know to do it." Small favors end with the question. Big favors never happen without the environment.
Imagine a new hire on day one. You say "please organize this" and nothing else. They don't know where the files are, what was decided in last week's meeting, who owns the project. They bring back the wrong thing. That person isn't stupid — the environment wasn't given. Humans have known this for decades. But somehow, in front of AI, that common sense disappears. Why? We were taught that AI only needs a question.
To make it concrete, let's use Claude's sub-agents as our example. Other AI tools in 2026 are moving in a similar direction.
When you open Claude Code, you get one chat window. In the old days, you'd type "write me a novel" and one AI would handle everything — genre, plot, characters — while the same window also fielded your off-topic "what should I have for dinner" comment. The whole context gets polluted. That's fine for a quick chat, but it breaks on project-scale work.
Sub-agents change this. If you're writing a book, you create a "book writer agent." It has its own context window — a separate room from the main conversation. You can chat about anything in the main window; the book writer just keeps working in its own space. You split the environment.
Think of a special operations team. When a mission comes in, you don't assign everything to one person. You assign a radio operator, a medic, a sniper, a demolition expert. Each one carries their own specialty and their own gear. The commander holds the whole plan, but actual hands and feet work inside each operator's own environment.
That's exactly the AI agent structure.
Giving everything to one person is like sending one unarmed operator into a contaminated site alone. Looks fine for a minute. Breaks fast.
Now the numbers. A single chat window typically holds around 200,000 tokens. A token, roughly, is a word and a half's worth of characters. A Korean sentence is about 30-50 tokens; one page of a book, 600-800 tokens.
So one window holds about 200-300 pages before it fills. If you write a whole book in one conversation, by chapter 3 or 4 it's already full. After that, AI forgets things. With sub-agents, each agent has its own window. Spin up 4 agents and your effective capacity is 4×.
Here's the aha:
AI isn't confused because it's dumb — its desk is too small.
There are two ways to grow the desk. Write smarter questions (prompt engineering), or split the environment itself (context engineering). From here on, the big work is won in the second one.
How do you split? Just ask yourself one question.
"Can one person finish this at one desk?"
Two answers.
Cleaning up an email, translating a paragraph, summarizing one page. A single-line question is enough. Making an agent for this is over-engineering.
A book, a 20-page report, a web app, a month-long plan. Start by splitting agents from day one. By chapter, by feature, by phase.
Mnemonic: "One desk → prompt. Many desks → context."
Let's make it concrete. Task: write a 12-chapter book with AI.
"Write the book. Starting chapter 1."
Through chapter 3, smooth. By chapter 5, AI starts forgetting chapter 1's setup. By chapter 8, the protagonist's name is shaky. You paste previous chapters back in every time. Wasted time, wasted tokens.
1. Structure agent — owns the book's structure and consistency. Chapter order, character relationships, world rules. Saves to file.
2. Writing agent — writes one chapter at a time. Opens the structure file at the start, then focuses only on the current chapter.
3. Review agent — reads a finished chapter, compares it to the structure file, points out drift.
With this split, each agent's desk stays clean. The structure agent knows all 12 chapters but doesn't write. The writing agent sees only the current chapter, and writes it deeply. The review agent sees only the finished draft and judges. Like a special ops team.
How do you actually make an agent?
In Claude Code, open Claude in the terminal, then type:
/agents
A management menu appears. You'll see options:
/agents # open agent management menu
> Create new agent
> Project level # use only in this project
> Personal level # use in all projects
For your first try, pick "let AI build it" and describe what you need in plain language. Something like "make me an agent that manages the structure of a book."
Do this for a week. Every time a task comes up, ask yourself: "one desk or many desks?" Awkward at first. Automatic in a few days. You'll watch the "AI got confused" moments drop off fast.
To recap.
Work has a question and an environment. The environment must fit for the question to matter. We explained the principle through Claude's sub-agents, but it applies to every AI tool and every future technology. The word changes. The skill of splitting environments stays central.
Carry one question — "Can this finish at one desk?" That single question separates prompt from context and auto-scales your approach. Any big task, split into agents. Each one minds its own desk.
People who write good prompts will become common. People with the instinct to set the table will last. That instinct doesn't come from AI — it comes from the oldest principle in work: division of labor. Two or three years from now, when "sub-agent" is gone as a word, today's principle still runs. Technology changes. Environment design doesn't.
Ask small. Set wide. Many desks.