Build a Multi-Agent System (From Scratch)

Multi-agent systems are among the most discussed and worked-on topics in AI today, but most engineers reach for frameworks like CrewAI or LangGraph without understanding what happens under the hood. This book takes you under the hood by having you build your own LLM agent framework—called llm-agents-from-scratch—from the ground up. Starting with the core insight that LLMs can only express intent but cannot act, you'll build the orchestration infrastructure that turns intentions into actions: interfaces for tools and LLMs, a processing loop, memory systems, and support for the core protocols MCP, Agent Skills, and A2A. Assuming basic familiarity with LLMs and Python, you'll code along chapter by chapter, gaining the deep understanding needed to work confidently with any existing framework or to build specialized solutions of your own. By the end, you'll be able to assemble focused agents into collaborating multi-agent systems and know exactly why and how each component works.

What it argues

A framework model expressing how design levers (backbone LLM capabilities, equipped tools, protocol integrations, processing loop design, and multi-agent composition) drive psychological/behavioral states of the agent system (planning quality, tool-use behavior, reusable-workflow adherence) that in turn produce task performance outcomes such as accuracy, efficiency, and versatility.

Key ideas it contributes

• Backbone LLM Capabilities — The planning and tool-calling competencies of the LLM at the core of an agent, determining how well it formulates plans and generates valid tool-call requests.
• Equipped Tools — The collection of tools made available to the agent's backbone LLM to expand its actionable capabilities.
• Protocol Integration — The extent to which the agent adopts external standards (MCP, Agent Skills, A2A) to access tools, reusable workflows, and other agents.
• Processing Loop Design — The orchestration architecture that drives task execution through repeated planning and tool-calling sub-steps until completion.
• Agent Planning Quality — The degree to which the agent forms correct plans and adaptively revises them based on prior sub-step results.
• Tool-Use Behavior — The agent's pattern of correctly selecting tools, forming valid requests, and using returned results instead of relying on parametric knowledge.
• Multi-Agent Composition — The design of combining multiple specialized agents into a collaborating system for an overarching task.
• Task Decomposability — The extent to which a complex task can be split into subtasks amenable to focused agents.