SIXTEEN PAPERS. ONE EXPANDING QUESTION.
HARNESS
ATLAS
The model is only the beginning.
Take apart the machinery that turns a language model into an agent. Then learn to build the loop.
What survives the moment
A memory policy decides what to retain, what to retrieve, and what to place back in context.
STORE → RETRIEVE → REINSERT ↗Conceptual anatomy · layers illustrate responsibilities, not a hardware architecture
01 / THE GUIDED PATH
Learn the loop.
Then change it.
Eight chapters. Concrete failures, working mechanisms, and a question to test your understanding at every step.
ABOUT AN HOUR, AT YOUR PACEBeyond the answer
What turns a model into an agent?
Close the loop
What changes when the world answers back?
Remember deliberately
What should survive a full context window?
From mistakes to procedures
When does a correction become reusable?
Divide the work
Why treat a long input as an environment?
Change the machinery
What can improve without changing the model?
Keep becoming
What changes when improvement happens mid-run?
Earn the result
How do you know the progress is legitimate?