Someone found the agentic engineering books too long and wrote a short, phone-readable version himself
Gappy uses model queries to write technical notes and has released a short book on agentic engineering. It has a higher mathematical density than most material online, and commenters warn that models reviewing each other can easily make the same mistakes together.
Gappy (Giuseppe Paleologo) found the agentic engineering books on the market too long and wrote a short version himself. He put this technical note on Google Drive, free to read, with a very direct goal: short enough to finish browsing on a phone, and more mathematically dense than most material online.
Agentic engineering refers to engineering around AI systems that can plan, call tools, and iterate through multi-step tasks. The term has recently been wrapped in a thick layer by vendor blogs and long videos, and there is not much short material that actually makes it onto paper. In the post, Gappy says he is used to using Astra or Fable queries to “write” technical notes, most related to finance, with a few on AI. He likes F. Fleuret’s free book The Little Book of Deep Learning and wants to write agentic engineering along those lines: short, phone-readable, and straightforward.
He admits the material is not fully there yet, but it is already readable. The math is heavy because the book reflects his taste. He calls math the Great Compressor: only things that can be compressed are worth writing short. Another benefit is saving time: readers can flip through a few pages and know whether they can get value from it, without first investing hours. He says he doesn’t want to take up other people’s attention, but if this helps someone, the feeling of wasting time will be a little better.
The link is here:
Google Drive
The discussion section is more honest than the main text. Macci X says writing originally bundled two things: producing the notes, and judging whether they are worth keeping. Using model queries only replaces the first. Theory warns that two models happy to agree with each other reviewing each other will quickly converge on a shared error unless one of them is rewarded only for finding counterexamples. 生晓bai says something similar: debate between models is useful until both confidently inherit the same bad premise. Gappy replies that it is unlikely, as he has been editing the model’s output.
Zhangir Azerbayev thinks the material is not general enough and argues that an agent is a POMDP policy, and perhaps even R can be removed. Kuldeep Pisda says that short, mathematical material like Fleuret’s is exactly what agentic engineering lacks; right now the internet is mostly vendor blogs and hour-long videos. Someone also asks for the accompanying Python files and README, but the post doesn’t answer. Another person asks whether it was really handwritten and says it would take them months to read it once.
The most useful piece of feedback here may be the demystification of “writing a book with a model.” Models can help produce paragraphs, but judging which paragraphs are worth keeping and which premises were wrong from the start still has to be done by humans. The value of a short book lies exactly here: it cannot cover much, but it can expose the author’s taste and judgment as quickly as possible. If you can keep reading, continue; if not, close it. For a free note, that’s not a bad deal.
发布时间: 2026-11-16 04:28