An Illustrated Guide to AI Agents (Agents column)

While flying back from London to Moscow, I managed to read this. book I can recommend it to everyone. It is written by Jay Alammar and Maarten Grootendorst, the authors of the cool book Hands-On Large Language Models. told. The new agent book is as good as the previous one, and it’s still in the process of being written and only half the chapters have been written. It is already visible that the new book continues the topic of LLM, but already in the world of AI agents - that is, in the first book, readers can see how LLM works. (nearly 300 illustration), and in the new - how to assemble agent systems from LLM. That is, the first book answers the question "what is the model inside", and the new one - how to build memory, tools, planning and coordination around the model. In general, I recommend reading in this order to first understand the engine, and then the architecture of the application.

If we talk about the readiness of the book, now the main chapters are ready.

1. Introduction Why do you need an "agent" at all, how it differs from just an LLM call, where the chatbot ends and the system begins. This is a chapter for aligning terms and the overall architectural picture. 2. Reasoning LLMs What changes when the model is able not only to answer, but also to pass chains of reasoning / test-time reasoning. This is an important chapter so as not to confuse “smart response” with “model thinking.” 3. Memory It's one of the most practical parts: context, short-term vs long-term memory, context engineering, how the agent "remembers" and why memory quickly becomes an architectural rather than just an ML task. 4. Tool Usage, Learning, and Protocols This chapter is about tools, function calling, integrations, and protocols like MCP. That is, the moment when the LLM stops being just a text generator and starts doing something in the outside world. 5. Planning and Reflection - here we are talking about task decomposition, plan reassembly, self-critique and feedback loops. This is an area where agency is beginning to affect reliability and cost. 6. Multi-Agent Systems When one agent is not enough, how to divide roles between several, as well as how to coordinate agents. This is where the A2A protocol comes in. (I used to be. told MCP vs A2A protocol)

Anyway, the book came to me and I will be following the new chapters. I think that it will be useful for engineers and techlids, who need not see “another demo”, but to build a mental model in their head: from which blocks AI agents are generally assembled and where real engineering risks live in this design. While Hands-On Large Language Models was a good entry into the LLM theme, An Illustrated Guide to AI Agents looks like a good entry into agency architectures and multi-agent systems.

P.S. More detail It's available at system-design.space.

#Architecture #Software #AI #Engineering #ML #Data #SystemDesign #DistributedSystems