A book on design patterns and best practices for building agentic systems with LLMs. It combines theoretical foundations with hands-on implementation, moving from foundational reasoning patterns (CoT, ReAct) through tool use, orchestration, and multi-agent protocols (MCP, A2A) to evaluation and production infrastructure. All code uses PydanticAI.
Build a proof-of-concept agentic platform using established patterns and best practices -- not a full enterprise system, but one that teaches the architectural principles needed to design, implement, test, and operate AI agent systems that can evolve into production-ready solutions.
Software engineers and ML practitioners who want to build agentic systems. Familiarity with Python and basic LLM concepts is assumed.
The book is organized in two sections. The first covers the building blocks: reasoning patterns, tool use, context management, orchestration, retrieval, and the two inter-agent protocols (MCP and A2A). The second section puts those blocks together into production systems: evaluation, data connectors, execution infrastructure, user interfaces, and a complete agent built incrementally from simple to fully distributed.
| # | Chapter | Topic |
|---|---|---|
| 1 | Foundations | What agentic systems are, how they differ from traditional software, modularity and design principles |
| 2 | Core Patterns | Zero-shot, few-shot, CoT, ToT, ReAct, CodeAct, self-reflection, verification, planning, human-in-the-loop |
| 3 | Tools | Tool use, structured output, discovery, schemas, permissions, workspaces, intro to MCP |
| 4 | Orchestration & Control Flow | Workflows, graphs, delegation, hand-off, long-running tasks, event-driven agents |
| 5 | RAG | Embeddings, vector databases, document ingestion and retrieval, evaluation, attribution |
| 6 | Context & Memory | Prompt layering, context engineering, compression, token budgeting, write-back patterns |
| 7 | MCP | Model Context Protocol: architecture, tools, prompts, resources, sampling, transport |
| 8 | A2A | Agent-to-Agent protocol: discovery, tasks, message exchange, security |
| 9 | Skills, Sub-Agents & Tasks | Sub-agent delegation, skill packaging, task lifecycle, composition comparison |
| 10 | Evals | Deterministic testing, eval frameworks, custom evaluators, AI-powered quality analyzers |
| 11 | Data Sources & Connectors | SQL, OpenAPI, file, and vocabulary connectors; NL2SQL; private data guardrails |
| 12 | User Interface | Chainlit, AG-UI protocol, error propagation, session identity, file uploads |
| 13 | Execution Infrastructure | Sandbox, REPL, MCP server isolation, skill sandboxing |
| 14 | The Complete Agent | Five progressive agent variants, then decomposition into distributed MCP/A2A services |
Full table of contents with section-level detail: chapters.md
API and module documentation for the agentic_patterns library: docs/agentic_patterns.md
chapters/ Book chapters (markdown)
agentic_patterns/ Python code
core/ Reusable infrastructure
agents/ Domain-specific agents
toolkits/ Business logic (no framework dependency)
tools/ PydanticAI tool wrappers
mcp/ MCP server wrappers
examples/ Code examples by chapter
testing/ Testing utilities
prompts/ Prompt templates
tests/ Unit and integration tests
scripts/ Build, validation, lint scripts
docs/ Reference documentation
output/ Generated book (book.md, book.pdf)
uv pip install -e .scripts/make.sh # generates output/book.md and output/book.pdfscripts/test.sh # runs unit + integration testsThe doctors command provides AI-powered analysis for prompts, tools, MCP servers, A2A agent cards, and skills.
doctors prompt prompts/system.md
doctors tool my_module:my_tools
doctors mcp --url http://localhost:8000/mcp
doctors a2a http://localhost:8001/.well-known/agent.json
doctors skill path/to/skill/Run doctors --help for all options.