This repository contains structured exercises designed to demonstrate the shift from "Autocomplete Coding" to "Agentic Engineering" using Claude Code.
In this repo, we do not treat the AI as a chatbot. We treat it as a Agentic Coding Partner. Every exercise is designed to fail initially, requiring the Agent to use Plan Mode to do the following:
- Analyze logs and cross-file dependencies.
- Propose a strategic plan.
- Execute only after human architectural approval.
We have shifted from simple "assistance" to full "engineering partnership." This is how the workflow has evolved over the years:
- 1990s – 2020 | Statistical Era (IntelliSense): Simple pattern matching. It finishes variable names but lacks any understanding of logic.
- 2021 – 2023 | Predictive Era (Autocomplete): Predicting the next line of code. Fast, but context-blind, often leading to "correct-looking" bugs.
- 2023 – 2024 | Junior Era (Simple Chat): Helpful for snippets and explanations, but impulsive. It focuses on the immediate chat request while skipping the system-wide plan.
- 2025 – 2026 | Senior Era (Agentic Partner): System-wide reasoning using Plan Mode. It analyzes the whole project and proposes a strategy before touching a single file.
Plan Mode decouples Thinking from Doing, shifting the AI from an impulsive "Chatbot" to a strategic Engineering Partner.
- Measure Twice, Cut Once: Scans the entire repo to identify root causes before touching a single line of code.
- Human as Architect: You audit a technical blueprint rather than cleaning up an AI-generated mess.
- Context Discovery: Eliminates "guessing" by forcing the agent to explore file dependencies and project structure first.
- Efficiency: 1 minute of planning prevents 20 minutes of reverting a broken build.
| Aspect | Autocomplete (Old) | Plan Mode (Agentic) |
|---|---|---|
| Logic | Impulsive (Guess & Write) | Strategic (Analyze & Propose) |
| Human Role | Janitor (Cleaning up) | Architect (Directing) |
| Outcome | Trial and Error | Right-First-Time |
The curriculum is designed to move from local compiler issues to complex system-wide architectural failures.
| Difficulty | Exercise Name | Stack | Core Learning Objective |
|---|---|---|---|
| ⭐ | rust-traits-mismatch | Rust | Fixing cross-file interface discrepancies. |
| ⭐ | docker-permissions | Docker, Node.js | Fixing Docker container permission errors. |
| ⭐ | middleware-ordering | Node.js, Express | Fixing Express middleware execution order bugs. |
| ⭐⭐ | rust-deadlocks | Rust | Resolving multi-threaded deadlock through locking order analysis. |
| ⭐⭐⭐ | distributed-race-condition | Docker, Node.js, Redis | Fixing inventory overselling across distributed API replicas. |
| ⭐⭐⭐ | rust-memory-leaks | Rust | Identifying and fixing reference cycles that bypass the borrow checker. |
| ⭐⭐⭐⭐ | k8s-mismatch | Kubernetes, Kind, Node.js | Debugging multi-layer Kubernetes networking and probe failures. |
- 2026-01-18: GitHub Copilot now supports OpenCode
- 2026-01-16: Ollama: Claude Code with Anthropic API compatibility
- 2026-01-14: GitHub Copilot CLI: Enhanced agents, context management, and new ways to install
- 2026-01-12: Even Linus Torvalds is vibe coding now
- 2026-01-10: Code Is Cheap Now. Software Isn’t.
- 2026-01-09: OpenCode no longer works with the Claude Max plan
- 2025-12-27: Andrej Karpathy on Vibe Coding