The Tutor That Measures Where You Stop Understanding

LearnHarness calls itself an agent harness for learning. The phrase is accurate and unhelpful, so the repository's own description is better: it uses AI agents to plan and teach lessons from your current understanding to your goal, and it builds a detailed picture of what you understand and how you reason.
The operative word is current. Most AI tutors begin at the beginning of a topic and explain forward. This one begins by finding out where you actually are, which turns out to be a different product.
Four stages, in a fixed order
Measure comes first. Before teaching anything, the agent probes each part the lesson will depend on, looking for the edge of your understanding. Answers, confidence and reasoning all count as signal. The skip button exists, but skipping removes the thing that makes the rest work.
Plan comes second. From that starting point it builds a dependency map to your goal. You can see the plan, and the plan adapts when the diagnosis turns out wrong. That adaptability is the part that separates a real map from a syllabus.
Teach is third, and its rule is the interesting one: each step has to build on something you already understand. You predict, then get an explanation, then check your understanding. Confusions get repaired before anything is stacked on top of them.
Practice is last and separate. Sessions pose varied problems that require the right knowledge to come to mind without being named first. Hints adapt to your reasoning rather than to a difficulty setting.
The stated aim is more precise than most education copy: a learner should end up finding the goal obviously true, rather than merely familiar.
The design bet
Every tutoring product makes a bet about what the bottleneck is. Most bet on explanation quality, so they invest in better answers. LearnHarness bets on diagnosis, which means it spends the user's first ten minutes on questions instead of answers.
That is an expensive bet. People who open a tutor usually want the thing explained, and a product that responds with a diagnostic feels like a delay. The counter-argument is that explanation is now nearly free and therefore no longer the constraint. Anyone can get a correct explanation of the chain rule in two seconds. What nobody can get cheaply is a correct diagnosis of why the chain rule stopped making sense at the third step, which is usually some earlier thing about composition that was never solid.
A dependency map is the mechanism for finding that earlier thing. It also explains why the plan has to be visible and revisable: if the diagnosis is wrong, the map is wrong, and the user is the only one who can tell.

Why the persistence matters
The knowledge graph carries across sessions. In isolation that sounds like a feature bullet. In practice it changes what the product can be, because a tutor that remembers is a tutor that can avoid re-teaching and can deliberately revisit.
It also creates the hardest problem in the category. A model's context does not persist on its own. Persistence means something is storing a structured claim about what you know, and that claim is now a document about a person. The repository does not dwell on this. Anyone building for schools will have to.
The desktop requirement is connected. The app accepts files, PDFs, photos, anything you can drag in or attach with the paperclip, and the file stays local rather than being uploaded to a service. That is why it ships as a desktop application instead of a web page. Whether it is a good trade depends on whether you think a student's textbook scans belong on someone's server.
The input also includes dictation and free-text thinking fields, where you explain your reasoning, uncertainties and reactions in your own words. That is the mechanism behind the "how you reason" half of the pitch. Without it the agent only sees right and wrong, which is the least informative signal a tutor can have.
The economics are the reason this is possible
Two years ago, a product that spent the first ten minutes of every session on diagnostic questions would have been commercially impossible. Inference was expensive, and curiosity was the thing you charged for.
Now the marginal cost of a question is close to zero, and the constraint moved. What is scarce is the structure that turns a conversation into a sequence with a beginning and an end, which comes from a design decision about what to ask and when to stop asking rather than from any model capability.
That shift explains the flood of learning tools in the trending lists this month. The models got good enough to hold a diagnosis, and the harness, the part that decides the order of operations and remembers what happened, became the product. LearnHarness's four stages are a position statement about where the value sits, and the position is that it sits in the sequence rather than in the answers.
What it does not solve
No outcome data. The repository describes a method, not a study. There is no comparison against a baseline, no retention measurement, no evidence that a diagnosed lesson beats a well-written explanation for a motivated adult. Education technology has a long history of plausible mechanisms and disappointing results, and this one arrives with the same absence of proof as everything else in the category.
Diagnosis quality is bounded by the model. If the agent's questions are poorly chosen, the map is built on noise and the entire session inherits the error. The repair step, where confusions get addressed before moving on, is exactly where a weak model degrades into reassurance.
The four-stage structure can also feel like friction to an expert. Someone who knows what they are missing wants the specific missing piece, not a dependency map. The product's answer would presumably be that experts aren't the target, which is a fine answer that narrows the market.
The thing worth copying
Strip away the branding and there is one transferable idea: put the measurement before the instruction, and make the measurement force the learner to produce reasoning rather than select an answer.
That works outside tutoring. Code review tools that ask what you expected before showing the diff. Onboarding that asks a new hire to predict how a system behaves before documenting it. Sales enablement that tests the objection you cannot handle rather than the pitch you can. In each case the expensive part is knowing where the gap is, and in each case the default design skips straight to the material.
LearnHarness is at 9,200 stars, which for a desktop education app with no free tier marketing is a strong signal that the diagnosis-first framing resonates. The method is old. Putting it in front of an agent that can run the diagnostic, hold the map across sessions and adapt the practice is new enough to be worth a look.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.