Principles 2Nd Edition Summary

Principles of Building AI Agents (2nd Edition)

Author: Sam Bhagwat | Published: May 2025 (2nd Edition) | Source: sources/principles_2nd_edition_updated.pdf

Overview

Written by the founder of Mastra (open-source JavaScript agent framework) and co-founder of Gatsby, this book is a practitioner's guide to building AI agents without hype. The 2nd edition (May 2025) adds new chapters on MCP, image generation, voice, A2A, web browsing/computer use, workflow streaming, code generation, agentic RAG, and deployment.

Key Themes

1. Agents Are Just LLMs Using Tools in a Loop

The book demystifies agents: they are not a magical new paradigm but a composition of primitives — LLMs + retrieval + tools + memory. Agency exists on a spectrum:

  • Low autonomy: binary choices in a decision tree
  • Medium autonomy: memory + tool calls + retry
  • High autonomy: planning + task decomposition + queue management

2. Tool Design Is the Most Important Step

"Think like an analyst. Break your problem into clear, reusable operations. Write each as a tool."

The book gives a detailed case study: Alana's book recommendation agent. First attempt (drop all books into context) failed. Breaking the problem into specific tools (by investor, by genre, by recommendation type) succeeded. Key insight: how you decompose operations into tools determines agent capability.

3. Memory Is Hierarchical

Memory systems combine:

  • Working memory: recent messages (sliding window via lastMessages)
  • Long-term memory: semantic recall via RAG (semanticRecall)
  • Memory processors: TokenLimiter (prune by token budget), ToolCallFilter (remove verbose tool calls)

4. MCP = USB-C for AI

Anthropic's Model Context Protocol (MCP, proposed Nov 2024) standardizes tool definitions. Hit critical mass March 2025 after Shopify CEO advocacy. April 2025: OpenAI + Google Gemini announced MCP support, making it a de facto standard. Similar to USB-C: protocol standardizes, ecosystem provides adapters. Open-source registries: Smithery, PulseMCP, mcp.run.

5. A2A for Multi-Agent Communication

Google's Agent-to-Agent protocol (A2A, spring 2025) handles communication with "untrusted" external agents. Uses /.well-known/agent.json for discovery, JSON-RPC 2.0 over HTTP, SSE for streaming. Still early; OpenAI/Anthropic haven't adopted.

6. Workflows > Pure Agents When Predictability Matters

Graph-based workflows add structure where agents alone are too unpredictable. Primitives: branching (parallel LLM calls), chaining (sequential), merging (converge), conditions (decision-based). Best practice: one LLM call per step maximum.

7. RAG Alternatives Exist — Start Simple

The book's controversial take: full context loading (Gemini 2M token window) should be the first approach. Only build RAG pipelines if that fails. Then try agentic RAG (tools over docs). Then full RAG pipeline with chunking/embedding/reranking.

8. Evals Are CI for AI

Unlike binary pass/fail tests, evals return 0-1 scores. Categories:

  • Textual: hallucination, faithfulness, completeness, answer relevancy
  • Classification: sentiment, topics, entity extraction
  • Agent tool usage: did the agent call the right tools?
  • A/B testing: live experiments with real users (Perplexity, Replit's primary quality signal)

9. Multimodal: Ghibli Moment Happened

March 2025 "Ghibli-core" moment showed consumer-grade image generation crossed into viral cultural phenomenon. Video generation hasn't reached that milestone yet. Voice: STT→LLM→TTS pipeline still dominant; realtime voice-to-voice models struggle with turn-taking (voice activity detection).

10. Deployment Is Still Heroku-Era

Most teams: agent in Docker → deploy to auto-scaling container service. Serverless not ready for long-running agent workloads (timeout issues, bundle size). Teams using EC2/DigitalOcean with B2B use cases (predictable traffic) sleep best.

Key Contrasts with Other Wiki Sources

TopicThis BookOther Sources
Agent definitionLLM + tools/memory loopVaries widely
Memory approachHierarchical (working + long-term)MemPalace: verbatim; Supermemory: LLM extraction
RAGLast resort after context loadingVaries
MCPUSB-C analogy, standards storyAnthropic paper focuses on protocol mechanics
Multi-agentSupervisor pattern, workflows as toolsVarious frameworks

Notable Quotes

"You should think of reasoning models as 'report generators' — you need to give them lots of context up front via many-shot prompting." (Ch. 3)

"The most important thing you should do is think carefully about your tool design... Write this out clearly before you start coding." (Ch. 6)

"We're engineers. And engineers can over-engineer things. With RAG, you should fight that tendency. Start simple, check quality, get complex." (Ch. 20)

Relationship to Other Wiki Content

  • [[llm-agents]]: This book provides concrete implementation patterns (autonomy levels, tool design, memory) that extend the conceptual framework
  • [[augmented-llm]]: The book's "augmented LLM" (retrieval + tools + memory) directly maps to the Augmented LLM concept
  • [[agent-computer-interface]]: Tool design principles in Ch. 6 align with ACI poka-yoke philosophy
  • [[workflow-patterns]]: Graph-based workflows (Ch. 12-16) extend the 5 workflow patterns with branching/chaining/merging/conditions
  • MCP (Ch. 11): USB-C for AI; not yet a standalone concept page; see summary Ch. 11
  • MemPalace: verbatim vs LLM extraction memory — live tension in the field; no concept page yet

⚠️ Inference: The book's treatment of memory (LLM extraction) aligns more with Supermemory than MemPalace's verbatim approach. This creates a live tension in the field between "extraction loses context" (MemPalace) vs "versioning+grounding compensates" (Supermemory + this book).

Limitations Noted

  • MCP ecosystem: discovery fragmented, quality inconsistent, config schemas vary per provider
  • A2A: OpenAI/Anthropic haven't adopted; may not win standards war
  • Realtime voice: still no production-ready solution due to turn-taking problems
  • Video generation: hasn't crossed "Ghibli moment" threshold
  • Security: MCP server vulnerabilities already demonstrated (GitHub MCP server leak); expect more incidents as agent deployment 10x

End of summary.