agenticloops-ai/agentic-ai-engineering

What it solves

Reading blog posts about AI agents leaves you with theory, not skill. This repo exists because building agents is engineering, not magic. It gives engineers a hands-on path to understand what powers popular agents like Claude Code, Codex, and GitHub Copilot by building the core pieces yourself: the agent loop, the tool executor, the memory layer, and the eval harness — all from first principles, before any framework is introduced.

How it works

The repo is a structured sequence of self-contained tutorials, organized into modules that progress from basics to advanced. You start with a single LLM call and build up: prompt engineering, chat, tool use, the agent loop, then more sophisticated patterns like routing, parallelization, orchestrator-workers, and evaluator-optimizer. Later modules cover advanced techniques (structured output, streaming, context engineering, cost optimization, memory, RAG, multimodal, guardrails), testing and evaluation (unit tests, evals, tracing, red teaming, benchmarking, eval harnesses), loop engineering (skills, hooks, sandboxing, MCP, subagents, compaction), frameworks (building the same agent nine different ways to compare trade-offs), and production concerns (12-factor agents, deployment, monitoring, security). Each tutorial is a Python script runnable with a single command (using uv), and many modules end with a capstone project that combines everything into a production-style agent.

Who it's for

Engineers with basic Python knowledge and curiosity — no prior AI/ML experience required. If you want to truly understand how agents work under the hood, answer whiteboard questions about tool calls and inferencing loops, or prepare for 2026 software engineering interviews where agent fluency is becoming baseline knowledge, this repo is for you. It's also useful for anyone who wants to build their own agent this weekend instead of reading 100 blog posts.

Highlights

• Build the agent loop, tool executor, memory layer, and eval harness from scratch — no black boxes, everything visible
• Runnable with one command: uv run --directory <tutorial> python <script>.py — no conda or Jupyter headaches
• Companion Substack series “Disassembling AI Agents” reverse-engineers real production agents (Claude Code, GitHub Copilot, OpenCode) and pairs with the tutorials
• Covers the full path from a first LLM call to a production eval harness, including advanced patterns, safety, and deployment
• Six planned modules (some marked “coming soon”) covering foundations, effective agent patterns, advanced techniques, testing & evaluation, loop engineering, frameworks, and production
• MIT licensed and backed by a community discussion and issue tracker