Introduction
This series teaches how to build a coding agent using LangGraph. I called it Axon, a terminal-native coding agent like Claude Code, Manus, Codex, etc.
The aim is to make end to end coding agent design: state, nodes, edges, tool-calling mechanism, the ReAct loop, planning, sub-agent orchestration, memory, self-correction. Each chapter implements one of these as a working feature.
What we build: Axon#
Axon is a CLI coding agent. By the end of the series it will:
- Use tools such as: read, write, edit, and search files in any repo.
- Ask for confirmation before destructive changes.
- Plan multi-step tasks and execute them.
- Spawn sub-agents that work in parallel.
- Persist project-scoped memory across sessions.
- Self-correct on failed tool calls.
- Run on any consumer or a local model
Each chapter ships a runnable feature on top of a real codebase, not a notebook toy.
This is the machine we end up with a few chapters in. Click around; by chapter 3 you will have built every box yourself:
Why build it from scratch?#
It's about maximum control over the state. You can use LangGraph's create_deep_agent or create_agent helper for a standard agent,
but if you need your own logic inside the graph custom nodes, you need to build from scratch.
LangGraph is a low-level orchestration framework and runtime for building, managing, and deploying long-running, stateful agents so it allows us to maximum control for our domain specific agent.
The stack#
- LangGraph: state machine, everything we need to build AI agents
- LM Studio: local LLMs to serve
Chapter map#
| # | Chapter | What Axon gains |
|---|---|---|
| 0 | Introduction (this post) | Series scope, rules, stack |
| 1 | State & Graph | Minimal REPL, the agent skeleton |
| 2 | Tools & ReAct Loop | Filesystem read, write, search |
| 3 | Human-in-the-Loop | Approval question for destructive operations |
| 4 | Planning | Multi-step tasks via plan-and-execute |
| 5 | Sub-agents & Orchestration | Parallel research and multi-file edits |
| 6 | Memory | Long-term project-scoped memory across sessions |
| 7 | Self-correction | Reflexion + critic loop |
| 8 | Evaluation | Trajectory eval harness, cost dashboard |
Up next: State & Graph#
Chapter 1 builds the smallest LangGraph that compiles: one node, one edge, a typed config, a working REPL. The four primitives introduced there - State, Reducer, Node, Edge - are what every later chapter extends.