You open Codex. Fifteen minutes later the weekly quota is gone. The job never finished.
Meanwhile, one developer in OpenAI's own rollout thread is not complaining. Same model, same limits, same week. The difference is not talent. It is that he owns the loop his agent runs in, and you are renting yours.
This book hands you the seven parts that decide whether a job finishes, and makes every one of them yours. By Chapter 4 you will have run one real job on two different models and written down four numbers for each. By Chapter 5 the model behind your agent changes with a single line. By Chapter 7 a long run stops growing its own bill. By Chapter 8 your setup catches the agent saying it finished when it did not. By Chapter 10 you have an evaluation set built from your own past work. By day 30 a new model launch costs you one morning instead of one rebuild.
Here is what you will build:
1. A model adapter that changes providers with one line of config.
2. A token budget with a ceiling you set and an eviction rule that holds it.
3. A state store that keeps run history out of the conversation, so a long job stops paying for its own memory.
4. A claim checker that tests the exit code, the bytes and the diff instead of believing the summary.
5. A routing table priced per finished job, with the failed reruns counted in, worth the $200 a month anyone is burning on work that never lands.
6. An evaluation set built from your own past work that runs against any new model in one command.
7. A stop that halts work at the tool boundary instead of asking the model to please stop.
You also leave with two things that are not code. A written inventory naming all six layers of your current setup and who controls each one, and the switching cost of your own stack in hours, so leaving is a decision instead of a fear.
This is roughly 180 pages of build steps you run on your own machine, on whichever model you happen to be paying for. Every part you build ships in a free companion repository, at the paths the book uses, so nothing here is retyped out of a book: the runner, the adapter, the meter, the store, both gates, the pricer and the eval set, plus an offline test suite and an installer that wires the hooks for you. Scroll up and grab your copy.
