Reverse Engineer Claude Code
Stop using agent frameworks you do not understand. Build a production-grade coding agent from scratch in 14 hours.
What You Will Learn
- ✓Understand the agent loop pattern that powers Claude Code and similar tools
- ✓Design safe, composable tools with path sandboxing and error boundaries
- ✓Implement three-layer context compaction for unlimited conversation length
- ✓Build file-based task management with dependency graphs
- ✓Create multi-agent teams with message buses and coordination protocols
- ✓Add git worktree isolation for safe parallel execution
Who Is This For
- AI engineers — You use agent frameworks but could not build one from scratch. That gap is becoming a liability.
- Backend developers — You want to understand how tools like Claude Code, Cursor, and Copilot actually work under the hood.
- Technical leads — Your team ships AI features but nobody understands the agent architecture. You need someone who can debug and extend it.
Curriculum
The Agent Loop
3 lessons
Tool Design
3 lessons
Planning and Constraints
3 lessons
Subagents and Skills
4 lessons
Memory Management
3 lessons
Parallelism
3 lessons
Team Coordination
4 lessons
Worktree Isolation
3 lessons
Frequently Asked Questions
- Do I need Claude Code installed?
- No. You build your own from scratch. The workshop teaches the patterns, not the product.
- What LLM provider do I need?
- Any provider that supports function calling. LiteLLM handles OpenAI, Anthropic, Google AI Studio (free tier), and more.
- Should I take Building AI Agents first?
- Yes. This workshop assumes you understand tool calling and agent loops. Building AI Agents covers those foundations.
- How long does the workshop take?
- About 14 hours. Most people finish in 3-4 sessions over a week. Each module is self-contained.
Reverse Engineer Claude Code
Reverse-engineer Claude Code. Build a production AI coding agent from scratch in 12 modules.
Chosen by Senior Engineers at Top Product Companies
Be honest. Which one hurts the most?
You use frameworks blindly
You can write prompts using LangChain or AutoGen, but when the abstraction leaks or breaks, you have no idea how to fix it.
Your agents are toy demos
They work on 3-file side projects but crash on 300-file repositories because they lack context compaction and graph reasoning.
You don't understand tool sandboxing
Giving an LLM direct bash execution access terrifies you. You don't know how to build safe error boundaries.
You're falling behind on AI Architecture
Everyone is talking about agentic workflows and multi-agent teams, but you haven't built one that actually works reliably.
The secret isn't a better prompt.
It's a better architecture.
Senior engineers aren't better because they know more syntax. They're better because they've:
We don't teach you concepts. We make you build the engine.
Not a theoretical course.
It's an engine rebuild.
We tear down the most advanced coding agent in the world, look at how the pieces fit together, and assemble them from scratch. You write the code.
The Reverse-Engineering Protocol
12 progressive steps to a production system.
The Core Loop
You build the bare-minimum 17-line agent loop with tool execution. The foundation.
Context & Safety
You implement three-layer token compaction and filesystem sandboxing so your agent is immortal and safe.
Task & Memory
You add a Graph-based TodoManager and give the agent persistent skills to remember how to solve problems.
Multi-Agent Scalability
You deploy background daemons, JSONL message buses, and Git worktrees for safe, parallel agent teams.
From 17 Lines to a Production Coding Agent.
The Agent Loop
Build the core loop that powers every AI coding agent. Start with 17 lines, then add bash execution, error handling, and dual-mode operation.
Welcome and Setup
Clone the workshop repo, configure your environment, and run the test suite to verify everything works.
The 17-Line Secret
The core of every AI coding agent is a simple loop: prompt, tool call, result, repeat. Build it in 17 lines.
The Bash Agent
Add error handling, timeout protection, output truncation, and dual-mode operation (REPL + subagent).
Tool Design
Move beyond bash-only. Design read, write, and edit tools with path sandboxing. Learn why the model IS the agent.
From One Tool to Four
Why bash-only fails for code editing. Add read_file, write_file, and edit_file tools.
Path Safety and Sandboxing
Prevent the agent from escaping its workspace with safe_path() validation.
The Edit Tool
Why exact string matching beats regex for code editing. The model IS the agent.
Planning and Constraints
Plans get buried in conversation history. Add a TodoManager with constraints that keep the agent focused.
Context Fade
Why plans get lost as conversations grow. Introduce the TodoManager as external state.
Making Plans Visible
Build the TodoWrite tool with status tracking and nag reminders.
Constraints That Enable
Max 20 items, one in_progress, activeForm, ordering. Constraints make agents more capable, not less.
Subagents and Skills
Isolate context with subagents. Externalize knowledge with skills. Optimize costs with cache-preserving injection.
Context Isolation
Why one conversation for everything fails. Subagents get their own message history.
Agent Type Registry
Different tasks need different tool sets. Build an AGENT_TYPES registry.
Knowledge Externalization
Tools vs Skills. Build a SKILL.md standard and SkillLoader.
Cache Economics
Cache-preserving injection patterns. System prompt vs tool result cost trade-offs.
Memory Management
Build three-layer context compaction and file-based task graphs that survive compression.
The Context Overflow Problem
Token estimation, context window limits, and why the conversation is append-only.
Three-Layer Compression
Micro, auto, and manual compaction layers. Each handles a different scale.
State Outside Context
File-based DAG tasks with dependency resolution. State that survives compression.
Parallelism
Background execution with daemon threads, notification draining, and persistent teammates.
Background Execution
Blocking vs non-blocking. Daemon threads for fire-and-forget commands.
The Drain Pattern
Notification queue, thread safety, and draining before each LLM call.
Persistent Teammates
From disposable subagents to persistent teammates with TeammateManager.
Team Coordination
JSONL message buses, shutdown and approval protocols, WORK/IDLE lifecycle, and identity preservation.
The JSONL MessageBus
Append-only inboxes for inter-agent communication. Why JSONL over databases.
Shutdown and Approval Protocols
FSM patterns, request_id correlation, and graceful termination.
WORK/IDLE Lifecycle
Autonomous agents that find work themselves via idle polling and auto-claim.
Identity Preservation
Re-injecting identity after context compression. make_identity_block keeps agents self-aware.
Worktree Isolation
The final module. Git worktrees for directory-level isolation. Tasks as control plane, worktrees as execution plane.
Two-Plane Architecture
Tasks are the control plane. Worktrees are the execution plane. Separate concerns, safe parallelism.
Git Worktrees in Practice
Create, run, status, remove lifecycle. Task binding and scoped execution.
The Complete System
EventBus for observability, audit trails, and the full lifecycle. You built Claude Code.
Detailed Curriculum
See every system, every week, in detail before you decide.
Code is cheap. Agents are hard.
AI can write functions.
It can scaffold tests, write boilerplate, and fix typos.
AI cannot:
- Design context compaction algorithms
- Architect multi-agent communication buses
- Implement safe bash error boundaries
- Debug silent agent loop failures
That's what Senior AI Architects do. We train you to be one.
Learn From Someone Who Actually Does This
I am the Head of Engineering at Jobbatical (EU Tech), with 8+ years of leadership and 15+ years of total experience in the software industry.
"Most engineers are not blocked by ability, but by lack of real system ownership."
This accelerator exists to give you what most jobs never will.
Guest Sessions From Engineers at
Live sessions on System Design, Career Growth, and Interview Preparation.
Get Full Access
Invest in your career. It pays back 100x.
Frequently Asked Questions
Do I need Claude Code installed?
What LLM provider do I need?
Should I take Building AI Agents first?
How long does the workshop take?
14 hours from now:
Confidence comes from building it yourself.
Register for this workshop
Drop your details and we will send you joining instructions.