AI Tutor, Don't Lecture Yet: LearnHarness First Tests What You Don't Know

LearnHarness gives the AI tutor idea a twist: first test what you know, then decide what to teach. It is an open-source desktop app; the author, Eero Alvar, calls it an agent harness, which can be understood as wrapping a learning workflow and tools around a model. The goal is stated modestly: make understanding usable, focus energy on directly interacting with the material, and spend less time managing the learning process.

Most AI learning products treat the chat box as a classroom. You ask, it answers, with a confident tone, and whether it addresses the gap you actually have is left to luck. LearnHarness changes the starting point: it first probes your grasp of the concepts this lesson depends on, builds a mental-model graph, and then plans a minimal dependency tree from what you know to the goal. The author writes in the tweet that good explanations and good practice questions depend on what you already understand and how you understand it. The premise is not new, but few products build it into the workflow.

Teaching has three steps. Probe: probe the dependent concepts. Plan: plan the minimal dependency tree. Teach: start from what you understand and teach one reasoning step at a time. At each step, it asks you to predict the next step first, then explains, then checks understanding. Prediction and checking are used to expose errors and confusion on the spot, rather than letting problems roll forward. The goal is stated with some ambition: make each step, and ultimately the target, seem obviously true, not just familiar.

Practice is another track. Following a line of reasoning and being able to recall an idea when needed are two different abilities. Practice uses varied problems to broaden and strengthen retrieval cues, with difficulty set at a point where you really have to struggle but can still make progress. When you get stuck, the agent gives the smallest possible hint and tries not to think for you. That calibration is harder than simply giving the answer, and it reveals more about the product's craft.

The shared record is the skeleton of the whole thing. Teaching sessions and imported notes are distilled into propositions and reasoning chains and stored as a graph. The graph contains ideas and premises, not ability scores. Future teaching can read the graph, and practice can draw material from it. System prompts are visible and editable, and learning records are plain local files. One commenter said plain local files are a quiet victory: later, when switching models, the graph can be re-derived without losing sessions. That judgment is sound.

The app itself is free and open source, under AGPL-3.0. You use your own key and model; there are no subscriptions or markups, and the interface has a real-time cost table. It supports 30+ providers, including OpenAI, Anthropic, Google, DeepSeek, and OpenRouter. For macOS Apple Silicon there is a signed and notarized DMG; the Windows x64 installer is currently unsigned and may trigger a SmartScreen warning; Linux has AppImage and deb. All three platforms support in-app updates. The author says that if it ever charges, it will only charge for convenience services such as hosted keys and hosted services, not for features that can run locally. Keep that on record and see what happens later.

The comment section is more interesting than the launch. Someone asked whether OAuth subscriptions could be connected; the author replied that ChatGPT subscription connection is in development and planned for a future version. Someone tried learning with a subagent and found that different implementations vary widely, was unsure which kind of tool was meant, and also ran into the project link not accepting web browsing. The author replied that research subagents currently need a search-engine key, and the barrier should be moved to the search tool so that scraping remains available. Someone asked whether learning is really this linear; the author said session dependency trees are rarely linear. Someone asked whether it is better for 40-minute focus sessions or 5-to-20-minute fragments; the author said they use both, depending on what is being learned. Someone asked whether it is ADHD-friendly; the author replied that it is more like a Socratic back-and-forth, not just reading. Two questions still had no public answer: what happens when the probe misjudges that you half-understand something, and can an incorrect proposition discovered in the graph be marked and its downstream dependencies tracked? The latter question is crucial: the more durable the shared record, the more important the correction path.

On the same day, another route surfaced in the comments. A commenter posted mmon.concept, a web platform that turns the knowledge dependency graph into a course medium. It asks users to first write a concrete goal, then evaluate through real tasks: for chess, seeing whether you can find a fork without hints; for writing a compiler, seeing whether you can independently handle unary minus; for drawing, seeing two-point perspective; for learning Portuguese, seeing whether you actively use the personal infinitive in conversation. Nodes light up based on demonstrated ability, not on finishing videos. It is free for two missions, $19 per month for unlimited, with models charged at cost or bring your own key. It and LearnHarness are desktop and web respectively, one leaning toward agent teaching and the other toward a task map, but their thinking collides on the same point: courses should follow a person's understanding.

After seeing all this, the judgment is simple. The core of an AI tutor is not how beautifully it explains, but whether it knows where you currently stand. 'Having explained it' is not the same as 'having learned it'; many products say this, but LearnHarness at least writes it into the workflow. Whether it can truly teach people depends on whether the probe is accurate, whether hints are restrained, and whether the correction path is smooth. Open source and local files give it one advantage: these things can all be inspected. Links are here: GitHub repo https://github.com/amosblomqvist/learnharness, download https://learnharness.ai. mmon.concept is here: https://mmon.co/ncept/. Original tweet: https://x.com/amosblomqvist/status/2113757368433393755.