Streaming chat REPL
A terminal interface with slash commands, safe interrupts and a live cost meter.
12-week mentorship · Python / TypeScriptBuild the AI coding agent
Build the AI coding agent
everyone else just uses.
A hands-on mentorship for engineers who are serious about AI. We start at the agent loop and finish with a real coding tool you built yourself — every subsystem, real code, no frameworks, no hand-waving.
- 12
- weeks
- 2
- tracks · py / ts
- 1
- real product
- 0
- frameworks
› fix the failing test in utils.ts⌕ grep_search "parseDuration" ✔ utils.ts:12 (1ms)◇ read_file utils.ts ✔ 16 lines [read-only]┌─ permission needed│ edit_file utils.ts│ - total += Number(value)│ + total += Number(value) * unit└─ [y]es · [a]lways · [N]o ▋
Week 5, your own tool, your own permission gate.
The agent loop✳Tool calling✳Evals✳Context engineering✳RAG✳Sandboxing✳Streaming✳Permissions✳Sub-agents✳No frameworks✳
Why this existsMost engineers use AI tools.
Most engineers use AI tools.
Very few can build one.
There is a widening gap in this field. On one side, developers who call an API and wire up a framework. On the other, engineers who understand what happens between the prompt and the answer.
People who can debug a runaway agent loop, explain why a tool call failed, cut a token bill in half, and prove a change made things better. That second group is small, and it is what companies are actually hiring for.
The fastest way in is not another tutorial. It is building the thing yourself, once, properly.
The premise. Build a working clone of a professional AI coding agent — the kind you already use every day — and you will understand this field better than almost anyone who only consumes it. That is the entire course.
The capstone
What you have by Week 12
Not notes. A tool — built one subsystem per week.
A real toolset
Read, write, edit, bash, glob and grep — behind a permission system you trust.
Semantic code search
A vector database you wrote yourself: chunking, embeddings, incremental indexing.
Context engineering
Assembled system prompts, project rules from AGENTS.md, and compaction that survives long sessions.
Planning & restraint
Plan mode, a maintained todo list, and a harness that decides what "done" means.
Durable memory
Resumable sessions, cross-project memory, and clean handoffs between days.
A secure sandbox
OS-enforced limits and a directory jail, so model-written code cannot hurt you.
Tracing & evals
Full request tracing, cost and latency observability, and a regression suite you own.
Multi-agent orchestration
Sub-agents with narrow toolsets, running in parallel and merging cleanly.
The roadmap
Twelve weeks, one real product
Three weeks of fundamentals most courses never reach, then nine that assemble them into a working clone of a professional coding agent.
Part I
Foundations
01Your First AgentThe loop that thinksShips: A working agent that reasons and calls tools02Evals & Engineering DisciplineMake it good, prove itShips: An agent you can measure and improve03The Runtime — Build a HarnessMake it survive productionShips: A runtime that runs any agent reliably
Part II
Build the clone
04The Chat InterfaceThe product beginsShips: mycode v0.1 — a streaming chat agent05Tools That Touch CodeGive the agent handsShips: A coding agent that edits your files — safely06Context & Prompt EngineeringOwn the context windowShips: An agent that reads your project's rules07RAG + Vector DatabasesGive the agent a mapShips: Semantic search over any codebase08Planning & Agent LoopsThink before you actShips: An agent that plans, then executes09Memory & SessionsNever lose your placeShips: An agent that remembers across days10Artifacts, Preview & SandboxSee it run, safelyShips: Live previews from a locked-down sandbox11Evals, Tracing & ReliabilityTrust, but verifyShips: An agent you can debug and trust12Multi-Agent OrchestrationA team of agentsShips: Your finished Claude Code clone — in your portfolio
How a week goes
Read it, build it, then read the real thing
- 01
Read the handout
A designed, self-contained handout — concepts, code, and the reasoning behind each decision.
- 02
Build it yourself
Every part is a runnable checkpoint. You type the code; nothing is copy-paste theatre.
- 03
Study the real thing
We open the actual Claude Code source and compare your subsystem with the production one.
- 04
Prove it works
Evals, tracing and measurement — so improvements are demonstrated, not claimed.
Fit
Who this is for
A good fit if you…
- Write Python or TypeScript comfortably
- Have used an AI coding tool and wondered how it works
- Want depth, not another API tutorial
- Are moving into AI engineering and need a portfolio project
- Prefer building the primitive over configuring the framework
Not a fit if you…
- Want to train or fine-tune models — this is systems, not research
- Are looking for no-code or prompt-only workflows
- Want a finished template to ship this weekend
- Cannot commit about six hours a week to building
instructorWho teaches it
Pouria Jahandideh
AI Engineer · Engineering Team Lead · 10 years in production
I build AI agents professionally — voice agents, an AI tax-preparation product, and ML tooling for medical imaging — and before that I spent a decade as a full-stack engineer and team lead across fintech, ride-hailing at 40M-user scale, and telecom e-commerce used in 10,000 stores.
I have been teaching that whole time: a full-stack curriculum of my own, and hundreds of students mentored through JavaScript, React, Python, Linux, DevOps and Kubernetes. This course is what I wish had existed when I started building agents.
Hiring
Roles worth a look
AI, ML and data positions I come across and think are worth sharing — written out in full on the board, so they stay readable after the original post disappears.
Questions
Frequently asked
Do I need machine learning experience?
No. This is systems engineering, not research. If you can write Python or TypeScript and have called an API, you have the background. We never train a model — we build the software around one.
Which language do I write in?
Either. Every week ships in Python and TypeScript, and the handouts carry both. Most people pick the one they already work in and read the other for contrast.
Why no frameworks?
Because a framework is exactly the thing you are trying to understand. You write the agent loop, the tool dispatch, the context assembly and the eval harness yourself — and afterwards you can read any framework and know what it is doing for you.
How much time per week?
About six hours: a live session, the handout, and the build. The build is the part that matters and the part nobody can do for you.
What do I actually have at the end?
A working AI coding agent you wrote — streaming REPL, real tools, permissions, memory, RAG, sandbox, evals and sub-agents — plus the portfolio repo and the ability to explain every line of it in an interview.
Is this cohort-based or self-paced?
Cohort-based, with live sessions and a group. The handouts stay yours afterwards, so the material is self-paced but the accountability is not.
Ready to build the thing
everyone else just uses?
By Friday of Week 12 you do not have notes. You have a tool — and the understanding to explain every line of it.
Message me on LinkedIn →Message me about timing, fit and pricing. I reply to everyone.