One Agent, Twelve Lessons, Zero Frameworks

What it does

This repository is a 12-lesson curriculum that constructs a single agent from the ground up using plain Python and a local GGUF model. Each lesson adds one capability—structured output, tool use, memory, planning, regression testing, telemetry—to the same evolving agent.py file. The goal is mechanical understanding rather than a drop-in library.

The interesting bit

The project treats agent construction as systems engineering rather than mysticism: no cloud APIs, no hidden chain-of-thought, and no personality theater. It explicitly rejects abstraction-heavy shortcuts in favor of constraints, state management, and observable loops.

Key highlights

• Runs fully offline against local GGUF models with no API keys or rate limits
• One continuously evolving Agent class that accumulates capabilities across all 12 lessons
• Covers the full lifecycle from basic prompting to regression testing (evals.py) and structured telemetry
• Philosophy doc defines agents as “loops, state, and constraints,” not thinking entities
• Isolated examples in complete_example.py let you inspect each lesson before combining them

Caveats

• The README warns this is for learning fundamentals, not production best practices
• The curriculum is strictly sequential; the authors tell you not to skip ahead
• Requires bringing your own GGUF model; the repository expects it in the models/ folder

Verdict

Worth bookmarking if you understand Python but feel lost inside abstraction-heavy agent frameworks, or if you want to see how state and constraints replace clever prompting. Skip it if you need a SaaS starter kit or expect hidden reasoning.