A Textbook for the 'I Can Call an API, Now What?' Phase of Agent Dev

What it does

agent_learning is an open-source, bilingual textbook—built as an mdBook—that walks through building LLM agents from first principles to production. It covers 23 chapters spanning tools, memory, planning, RAG, agentic reinforcement learning, multi-agent systems, evaluation, security, and deployment. The repository also ships a minimal, runnable reference-agent baseline with offline tests, an MCP server, and FastAPI endpoints, intended as readable teaching material rather than a framework.

The interesting bit

Instead of starting with framework APIs, the book insists on teaching the mechanism first—why an abstraction exists before how to import it. It backs this up with 330-plus original SVG diagrams, five interactive demos, and paper-to-practice notes on ReAct, Reflexion, MemGPT, GraphRAG, and GRPO. That is a lot of editorial discipline for a GitHub repo.

Key highlights

• 23 chapters in English and Chinese, totaling 188 Markdown pages per language
• 330+ original SVG diagrams and 5 interactive demos for architecture and training flows
• Runnable reference-agent with ReAct loop, memory, prompt-injection guardrails, and 16 offline tests
• Paper-to-practice coverage bridging research (Agentic RL, MCP, A2A) to engineering implementation
• Explicitly not an awesome-list or framework manual; structured as a connected mental model

Caveats

• The reference-agent is intentionally small and described as a “teaching baseline, not a claim of production completeness”
• Several roadmap items—searchable diagram gallery, more capstone implementations, evaluation starter templates—remain unchecked

Verdict

Worth bookmarking if you are tired of fragmented blog posts and want a coherent engineering narrative for agent construction. Skip it if you are looking for a drop-in framework or a curated link list.