We are having a problem with the new AI agent economy. It seems unsustainable for the amount of processing required, as our LLms are not very efficient machines.
We built an interactive explainer that makes the economics viscerally clear in your browser. Try it.
The Thing Nobody Talks About
Every time an AI agent takes a step, it doesn’t just read your latest message. It re-reads everything: your system prompt, the full conversation history, every tool call, every file it opened, etc. All of it and many times to create an output one predicted token a time. That’s how autoregressive models work.
See It for Yourself
The explainer lets you replay five scenarios and watch the token meter spin in real time:
- Chatbot chat — the cheap baseline, just to calibrate your intuition.
- Simple bug fix — a few tool calls, a couple of files. But getting expensive fast.
- Runaway loop — the agent gets stuck retrying. Things get difficult.
- Vibe-coding session — a realistic longer agent workflow. Running all over the world now many, many times over.
- Custom scenario — set your own parameters and see what breaks.
For each step, you see exactly what’s inside the context window: system prompt, messages, model “thinking”, tool calls, file contents. Then you watch the cumulative input token count, output tokens, and running dollar cost update live.