The learning graph is compelling, but it is still testing you

A developer tried the open-source AI learning tool LearnHarness on X, with the task of learning to build backpropagation from scratch. Let's put the conclusion up front: the real-time learning graph is very engaging, but the teaching part hasn't kept up. The post author's own words were that at times it felt like being tested rather than taught.

LearnHarness is positioned as an agent harness for learning. The project is open source under AGPL-3.0; chat history, the knowledge graph, settings, and API keys are all stored locally, and the model uses your own key or a compatible endpoint. The desktop app covers macOS, Windows, and Linux. It breaks learning into four steps: measurement, planning, teaching, and practice. It first probes the boundaries of understanding, then generates a dependency graph from the current starting point to the goal, then teaches step by step, and practice sessions draw material from the knowledge graph.

The trial task was to learn backpropagation from scratch. LearnHarness first asks probing questions, then uses Mermaid to generate a learning path on the spot. The path combines the learning content, goals, and starting-point assessment. When answering questions, the corresponding nodes are highlighted, so you can see which part of the path you are on and also why rates of change come before backpropagation.

That is also where the problem lies. The post author answered questions back to back and repeatedly received corrections; for some content they hadn't yet understood, they were asked to demonstrate understanding. At that moment, what was needed was explanations, diagrams, and worked examples, not another probing question. One responder, Tim Moon, asked how to make it more like a teacher: it should notice where the learner is stuck and help them get through it, rather than jumping straight to the next question. Robin's reply was more direct: the real-time graph provides visibility, but deep explanations are missing; the flow feels more like a quiz than a lesson. There was also a Chinese-language reply worrying that if probing questions keep coming one after another, learning might make people more and more anxious.

The knowledge graph is the more noteworthy part of LearnHarness. It stores what has been learned as propositions and reasoning chains, recording individual judgments and premises and linking them back to the source conversation. Subsequent teaching can read the graph instead of starting from scratch. Practice can also draw questions from the graph. It stores ideas and connections, not ability scores. This direction is closer to learning than simply recording right and wrong answers, but the premise is that the explanations on the nodes are sufficient; otherwise, the prettier the graph, the stronger the feeling of being tested.

The post author's next step is to move the working graph into an AI tutor whiteboard. When stuck at a node, calculate an example together, keep the calculation process visible, change one input, and trace which values change along with it. The author says they are building it. This direction isn't new; the hard part is connecting the graph with explanations so that nodes can expand into actionable steps and don't stop at 'you got it wrong.'

Before trying LearnHarness, the post author used Teachable Code to explore the GitHub repo, then read its teaching method. Teachable Code turns a codebase into an interactive course, keeps the system diagram visible, and lets you follow the current step to look at source code and inputs/outputs. 3D is used for tensors and memory; 2D is used for applications and tool calls. It provides five examples: mini-SGLang, nanoGPT, SWE-agent, DeepTutor, and AI chatbot. Each example can run in the browser and has Chinese and English pages.

There is a sentence in Teachable Code's documentation worth copying for all teaching tools: clearly state the boundaries of the evidence. The five courses demonstrate teaching formats; they do not reproduce the results published by the upstream projects. nanoGPT uses a self-built small CPU model to follow the source code structure and does not run upstream code; mini-SGLang is a replay of caching rules, and token counts are not a speed benchmark; SWE-agent uses recorded demonstration text submitted upstream and does not call a live model; the prompts and tool schemas for DeepTutor and AI chatbot come from the source code, and the conversations and tool results are only illustrative. The learning outcomes have not been evaluated with learners. This kind of candor is more useful than the phrase “smart tutor.”

Looking at the two projects together, you can see the same dividing line. Among the three things—measuring understanding, generating a path, and turning the path into teaching—the easiest is drawing the graph, and the hardest is laying out the reasoning when someone is stuck. LearnHarness's graph can tell you where you are; it can't walk the path for you. Open source and local storage make system prompts visible and editable, which suits developers who want to study agent harnesses; ordinary users have to configure models and handle keys themselves, so the barrier to entry is not low. If a learning tool ends up with only question generation and grading, it is just moving a workbook into a chat box.