Agentic Engineeringby Asem Abdo
0 of 8 lessons
openplanr.dev

Free course · for engineers who already use coding agents

From using agents to engineering them.

You already ship code with Claude Code, Codex or Cursor. This course gives you the discipline behind it: context engineering, harness engineering, loop engineering, orchestration and evals, in clear terms and with runnable examples. Every code sample is a small harness written in plain Python and was executed with the real interpreter when this page was built, so every output you see is real.

By · Updated

  • 8 lessons
  • 3 practice tools
  • 33 runnable samples
  • 105 drill items
  • 95 flashcards
Start with part 02No account needed.

The path

read in order, or jump to what you need

Foundations

  1. 02Agentic engineering, gently
  2. 03The agent loop and tools
  3. 04Context engineering

Context and harness

  1. 05Harness engineering
  2. 06The primitives, hands on

Loops and proof

  1. 07Loop engineering
  2. 08Multi-agent orchestration
  3. 09Proving it: evals, cost, security

Practice

  1. 10Lab
  2. 11Roadmap and sources
  3. 12Glossary
  4. 13Flashcards
Start with part 02

How to use it

Answer before you look

Every prediction hides the output until you commit. Wrong answers teach more than right ones.

Every output is real

Each sample ran with Python 3.14.8 when the course was built. Run shows what it printed.

Drill until it sticks

Labs, the error decoder and spaced flashcards turn the lessons into habits.

The map

Four disciplines, one stack

The vocabulary moved fast. In February 2026 Andrej Karpathy proposed agentic engineering for professional work where you orchestrate agents instead of writing most code yourself. Under that umbrella, practitioners now use three more precise terms. They nest:

Agentic engineering you orchestrate agents and oversee their work Loop engineering goal, checks, stop condition, budget Harness engineering tools, permissions, hooks, guides, sensors Context engineering what the model sees at each step Instructionsprompt and rules Toolsnames and schemas Memoryretrieved docs, notes Historyearlier turns Tool resultswhat came back Taskgoal and current state reads Model proposes the next action
Each layer contains the one inside it. The model reads only what the context window holds, so better context, harness and loops help more than better words.

The three mental models

Everything here comes back to these

Three rules that explain agentic engineering

  1. Agent = model + harness. The same model performs very differently depending on its instructions, tools, permissions and context. Most reliability comes from the system around the model.
  2. Enforce, don't request. A rule in a prompt is a request the model may forget. A rule in a hook, a schema, a test or a permission is a guarantee.
  3. The context window is a budget. Every token competes for attention and money. Load what the step needs, when it needs it, and nothing else.

Short on time

The 90-minute route

  1. Vocabulary and the loop

    Page 2 and page 3. 25 minutes.

  2. Context and harness

    Page 4 and page 5, the core. 30 minutes.

  3. Loop engineering

    Goals, checks, stop conditions: page 7. 15 minutes.

  4. Lab + cards

    L1 drills, then 20 flashcards. 20 minutes.

A course your coding agent can read too

Every page has a Markdown version

Point your agent at a page with .md, at llms.txt, or at the docs MCP server, and it reads the same lessons, labs and answers you do.

llms.txtMCP server

Your datano account, no server

Progress is saved in this browser as you go. Coming from the standalone course? Import its progress file.