Principles of Building AI Agents

Author: Sam Bhagwat Genre: Artificial Intelligence

About This Book

With large language models making “agent” style apps more common, this book offers a quick pass over the typical components of agent systems-models, prompts, tools, memory, retrieval, and evaluation. It’s more of a lightweight orientation than a deep guide, and the pacing reflects that: some sections are extremely brief (to the point where they can feel almost prompt-like), so you’re mostly getting a checklist-level tour rather than detailed instruction.

Key Insights

• Basic building blocks: Providers, models, prompts, tools, and memory show up as recurring primitives in most agent architectures-useful framing, even if it stays high level.
• Agent workflows: Planning and sub-tasking (planner/critic/executor patterns) are presented as reliability levers, but the treatment is more “what and why” than “how to implement well.”
• RAG basics: Retrieval-augmented generation gets a concise overview-chunking, indexing, and retrieval quality matter, though you won’t find much depth beyond the fundamentals.
• Observability & evals: Tracing and lightweight evaluation loops are highlighted as practical ways to reduce regressions and compare approaches, which is one of the more grounded parts of the book.
• Operational hygiene: Version prompts and tools, capture inputs/outputs, and prefer deterministic tool contracts to keep behaviour testable-solid advice, but not novel.

Recommendation

As a short intro, it’s perfectly serviceable: you’ll come away with the main vocabulary and the broad shape of modern agent architecture (tools, memory, workflows, RAG, tracing, evals). At the same time, it didn’t feel like it offered much you couldn’t get from a few good blog posts, and the frequent Mastra.ai angle makes parts of it read more like product positioning than a fully neutral guide. That said, I picked it up for free at the World Summit AI in Amsterdam (2025), and for a quick skim it’s still decent-just go in expecting a light overview rather than something comprehensive.