Token Economics
Before you reach for a compression tool, understand what it compresses. Three focused lessons: what a token is, why agent loops get expensive, and how to compress them without changing a single decision the model makes.
The one-sentence version. An LLM charges by the token. An agent re-sends its whole history every turn, so cost grows with the square of the conversation. Compression removes tokens the answer doesn't depend on — and the only honest kind is one you can prove didn't change the model's decision.
The three modules
~4 min
1
Fundamentals
What a token actually is, why structured text costs double, and where an agent quietly burns thousands of them before it reads your task.
- BPE & the anatomy of a token
- Prose vs. code vs. JSON density
- The agent-request token budget
Cost & Compression
Why multi-turn cost is quadratic, how caching rewrites the math, and every family of compression invented to fight back — with honest risk framing.
- The O(n²) multi-turn tax + a live calculator
- The compression family tree
- Why training a model helps — and how it fails
Distil & Proof
How to prove compression didn't hurt, how Distil applies everything in certified tiers, a practical playbook, and where the field is heading.
- Decision-equivalence & non-inferiority
- The certified tier ladder
- A best-practice playbook + SOTA
What you'll be able to do
- Read a token bill — know exactly which parts of an agent request you're paying for, and which are waste.
- Predict multi-turn cost — understand the quadratic, and why prompt caching is the single biggest lever.
- Choose a technique — match the right compression family to your risk tolerance, instead of guessing.
- Demand proof — tell genuine, certified savings from "it still looks fine."
New here? Take the modules in order. Know tokens cold? Jump to Module 2 for cost mechanics and the compression toolbox, or Module 3 for proof and the tier ladder — both cross-link to Concepts and Techniques for implementation depth.