The Roadmap

Twelve weeks,
one real product

Every week is implemented hands-on in real code, in your choice of Python or TypeScript. The first three weeks build the fundamentals every AI engineer needs; the next nine assemble them into a working clone of a professional AI coding agent.

Part I · Weeks 1–3

Foundations

The primitives of modern AI engineering: the agent loop, evaluation, and a production-grade runtime. Three weeks that most courses never reach.

Part II · Weeks 4–12

Build the Clone

One subsystem per week. Each week your tool — mycode — gains a real capability, and by Week 12 it reads codebases, edits files, plans, remembers, previews its work, traces itself, and orchestrates a team of sub-agents.

Week 4~6 hours, hands-on

The Chat Interface

The product begins

The capstone begins. Build a streaming terminal REPL with slash commands, safe interrupts and a live cost meter — the interface layer that makes an agent feel like a product.

Streaming terminal REPLSlash commandsSafe interruptsA live cost meterInside Claude Code: the real REPL
🛠 You ship: mycode v0.1 — a streaming chat agent
Week 5~6 hours, hands-on

Tools That Touch Code

Give the agent hands

A chatbot talks about your code; an agent changes it. Add the six tools that make a coding agent — read, write, edit, bash, glob, grep — plus the permission system that makes it safe to live with.

Read · Write · Edit · BashGlob & Grep searchThe permission systemTool error designFixing a real bug end-to-end
🛠 You ship: A coding agent that edits your files — safely
Week 6~6 hours, hands-on

Context & Prompt Engineering

Own the context window

The highest-leverage week of the course. Assemble the system prompt, read project rules from an AGENTS.md the agent discovers itself, and survive long conversations by compacting without losing what matters.

Assembling the system promptAGENTS.md project memoryToken budgetsCompaction for long sessions
🛠 You ship: An agent that reads your project's rules
Week 7~6 hours, hands-on

RAG + Vector Databases

Give the agent a map

Grep finds exact strings; it cannot answer "where do we handle rate limiting?" when the code says throttle. Build a vector database from scratch — chunking, embeddings, cosine similarity, incremental re-indexing — and wire it in as a tool.

Chunking strategiesEmbeddingsA vector store from scratchIncremental indexingGrounded codebase Q&A
🛠 You ship: Semantic search over any codebase
Week 8~6 hours, hands-on

Planning & Agent Loops

Think before you act

An agent with tools but no plan overreaches on some tasks and quits early on others. Add a read-only plan mode, a todo list the agent maintains, feature lists the harness verifies, and a maker/checker loop.

Plan modeTodo lists the agent maintainsFeature lists & scope controlLoop engineeringMaker/checker separation
🛠 You ship: An agent that plans, then executes
Week 9~6 hours, hands-on

Memory & Sessions

Never lose your place

Close the terminal and a normal agent forgets everything. Add resumable session transcripts, persistent cross-project memory, and a handoff file so a fresh session continues exactly where the last one stopped.

Resumable sessionsPersistent cross-session memoryWorkspace managementClean handoffs
🛠 You ship: An agent that remembers across days
Week 10~6 hours, hands-on

Artifacts, Preview & Sandbox

See it run, safely

Add artifact rendering and a live-preview server so generated pages refresh as the agent edits them, a background-task runner for long jobs, and a real OS-enforced sandbox so model-written code cannot touch what it should not.

Artifact renderingA live-preview serverHardened secure sandboxBackground tasks
🛠 You ship: Live previews from a locked-down sandbox
Week 11~6 hours, hands-on

Evals, Tracing & Reliability

Trust, but verify

Add full request tracing, cost and latency observability, an eval suite for your own agent, and the reliability patterns that let it recover from its own mistakes.

Full request tracingCost & latency observabilityAn eval suite for your agentReliability & self-healing
🛠 You ship: An agent you can debug and trust
Week 12~6 hours, hands-on

Multi-Agent Orchestration

A team of agents

Sub-agents, parallel orchestration and small focused agents — then final capstone assembly. By Friday you have a complete AI coding agent you built yourself.

Sub-agents & the Task toolParallel orchestrationSmall focused agentsCapstone assembly & showcase
🛠 You ship: Your finished Claude Code clone — in your portfolio

By Friday of Week 12, you don't have notes. You have a tool.

A complete, working AI coding agent you built from scratch — and the deep understanding to explain every line of it.

Message me on LinkedIn →