The method
This training is based on a pedagogical choice we want to make explicit from the start. We could have offered you a catalog of tools alongside installation recipes. We will not, because this kind of content goes stale within a few months: packages change names, configuration options evolve, and not much is left to take away a year later.
We take the opposite approach. The backbone of the training is the harness itself, that is, the set of functional building blocks it must include to work: context management, tools, delegation, orchestration, memory, safety, and verification. We first establish which building blocks are necessary and why, then we rebuild each of them by hand using open source software. Finally, we work our way up to the transferable principle, the one you will keep regardless of the tool of the moment.
The goal is not to build a competitor to Claude Code, and the reconstruction is deliberately minimal. What you take away at the end, rather than software, is the understanding needed to build your own harness, adapted to your uses, and to knowingly operate the harnesses you will use every day.
The triptych
Each reconstruction module unfolds in three steps that we repeat throughout the training.
The first step, Understand, starts from the need. What is the building block for, why is it essential, and how does a real harness implement it?
The second step, Reconstruct, consists in writing by hand, on Pi, the minimal equivalent of the building block. This is what puts the concept to the test. Hands-on implementation is always more effective than passive reading. We remind you that the code is an illustration, not the lesson. A harness is not a recipe that works for any use case. It feeds on your needs.
The third step, Generalize, brings out the principle that is independent of the tool used. It provides the design rules you would apply elsewhere. This is the step that truly counts, because it is the only one that never goes stale.
In addition to these three steps, at the start of each module you will find its objectives, the description of the deliverable when relevant, and references to explore the concepts further.
The structure
The training is organized into four acts.
| Act | Content |
|---|---|
| 1. Foundations | LLMs and their ecosystem, the building blocks of a harness, the Pi starter harness, and the method |
| 2. Brick-by-brick reconstruction | Context, tools, agents, workflows, memory, permissions |
| 3. Verify, evaluate, observe | Tests, multi-model evaluations, observability |
| 4. Building your own harness | A personal use case, and sorting durable / disposable |
We wanted this document to be as detailed as possible so you can experiment on your own.