Liwei Two Minutes: Token Economics in Plain Language #3 — Why Do Agents Suddenly Feel Human?
People used to think ChatGPT was already very human-like. It's not. Not even close.
Why? Because traditional chatbots are fundamentally "one question, one answer." You ask one thing, it replies once. Like a high-end customer service rep.
The real change happened when AI started "working on its own." That's the hottest thing right now: Agents.
The first time you play with an Agent, it's shocking. It suddenly acts like a real employee.
It breaks down tasks on its own, writes code, runs tests, reports errors, fixes bugs, keeps going. It even "talks to itself" while working.
Why this sudden change? The reason isn't mysterious. Because AI started burning its own tokens.
In the ChatGPT era, tokens mainly came from human input. You type some words, the model replies. The token flow was simple: Human → AI → Human.
Agent era is different. Now the token flow is: AI → AI → Tool → AI → AI. So tokens are burning inside the machine in loops.
Here's an example. Say you tell an Agent: "Build me a website."
A traditional chatbot would just give you a block of code. Done. But an Agent won't.
It will first analyze the task. Then start talking to itself: "Let's decide on the tech stack..." "Need React..." "Probably need a database..." "Generate the homepage first..." "Run the tests..." "Got an error..." "Fix and retry..."
Notice: this "thinking process" itself consumes tokens. And it consumes a massive amount.
Because the Agent isn't "generating the correct answer once." It's more like trial and error. Just like a human engineer: write, revise, test, redo.
So token consumption suddenly exploded. Before: user asks one question, AI answers once. Now: the AI might have run hundreds or thousands of token cycles internally. And humans only see the final result.
This is a lot like the Industrial Revolution. At first, coal was just for cooking. Then people discovered coal could power steam engines. And the entire industrial system started running itself.
Today's tokens are the same. Initially, tokens were just for chatting. Now they're driving the "internal thinking flow of machine work."
So a very strange new phenomenon has appeared in the AI world: Many tokens are no longer for humans to see. They're machine-to-machine communication.
In the future, human-generated tokens might only be a tiny fraction. The real token flood will come from AI-to-AI interactions. One Agent calling another Agent, one model orchestrating another model, a swarm of AIs collaborating on projects.
So the entire AI industry is starting to look more like an automated industrial system. No longer just chat software.
This is also why so many people have suddenly realized: AI is getting more expensive, more power-hungry, more dependent on data centers.
Because what's really being burned today isn't "chat content." It's the machines' own workflows.
In the internet era, humans uploaded information to the network. In the Agent era, humans are uploading "work" to AI. And tokens are the fuel that machine labor truly consumes in this new era.
