A private tutor that interviews you, builds you a real course from cited sources, teaches it, and then interrogates you to find out whether you actually understood any of it.
Anyone can generate a study plan now. That's the commoditised half. The part nobody does well is telling you, honestly, that you don't understand something you think you understand.
LRNR starts with an adaptive interview to work out what you actually know. Then it researches your topic from real fetched sources, with citations constrained to those sources, designs a session path, and teaches it one day at a time.
The agent I care about most is the evaluation one: it probes, and it presses when your answer is thin. Content is abundant and free. Honest verdicts are scarce.
Click any stage to read what it does, or press the button to walk the whole thing.
Every agent keeps a deterministic fallback, gated on test mode and on any model error.
Free-tier model quotas made this concrete fast. The pipeline makes ten to fifteen model calls per run, and free tiers rate-limit long before that. Early on, agents quietly degraded to templated output while still returning a successful-looking response. That is the worst possible failure: silent.
So every agent got a declared deterministic fallback, and the model layer had to report degradation rather than hide it. The test suite runs the entire pipeline offline with no network at all, which is only possible because those fallbacks exist.
Press Kill the model above and step through. Nothing crashes, nothing lies, and the evaluation agent refuses to issue a verdict it can't ground.