The Course That Completed Itself

Here is something that happened this month and deserves more attention than it received. An agentic AI tool called Einstein, built by a third-party developer, demonstrated that it could navigate the Canvas learning management system, complete assignments, participate in discussions, and finish entire courses without a human student doing any of the work. Not as a proof of concept. As a product.

Weeks later, Canvas parent company Instructure responded by unveiling its own agentic AI, designed not to complete courses but to assist instructors: automating rubric generation, aligning content to learning objectives, reviewing discussion threads. The framing was deliberate. Where Einstein replaced the student, Canvas positioned its tool as an amplifier for the teacher.

The press covered this as a rivalry. A battle for the soul of the LMS. But the real story is quieter and more unsettling. Both tools, the one that cheats and the one that helps, reveal the same structural truth about how we have built education technology. And that truth is worth sitting with.

The Architecture of Absence

Einstein was able to complete entire courses because the courses were completable by a machine. That is not a failure of AI security. It is a mirror held up to curriculum design. If every assessment is a deliverable, every discussion a checkbox, and every assignment a document to be submitted and graded, then of course an autonomous agent can do it. The system was not designed to detect thinking. It was designed to collect artifacts.

This is the uncomfortable question that Einstein forces into the open: how much of what we call "coursework" is actually just production? How many assignments measure whether a student can generate output rather than whether they grew in the process?

The answer, if we are honest, is most of them.

The Teacher Side of the Coin

Canvas's response is more thoughtful than it might first appear. By building an AI that handles what they call "low-value" tasks for faculty, they are acknowledging something important: teachers are spending enormous amounts of time on work that does not require their expertise. Rubric formatting. Content tagging. Discussion monitoring. These tasks are necessary, but they are not where the magic of teaching lives.

The magic lives in the moment a teacher notices a student has been staring at the same paragraph for ten minutes. In the decision to ask a different question when the standard one falls flat. In the conversation after class that changes the trajectory of a semester. No AI can do that. Not because the technology is insufficient, but because those moments require something AI does not have: a relationship built on accumulated trust.

So Canvas is right to automate the mechanical. The question is whether freeing teachers from busywork actually leads them toward more meaningful interaction, or whether the system simply generates new forms of busywork to fill the gap.

The Missing Layer

Here is what neither Einstein nor Canvas's teaching agent addresses: the space between the assignment and the submission. The revision. The pause. The false start that led to a real insight. The paragraph that was deleted and rewritten three times before it said what the student actually meant.

That space is where learning lives. And in most educational technology, that space is invisible.

Einstein can complete a course because courses are built around endpoints. Canvas can help teachers manage those endpoints more efficiently. But neither tool makes the journey between endpoints visible. Neither captures the process of thinking itself.

This is the work we are pursuing at Koan. Our AI tutor, Aidan, does not complete assignments for students. It does not even draft them. It asks questions. Socratic questions calibrated to the student's rubric and adapted to their patterns. And every interaction, every revision, every pause between keystrokes is captured in the student's learning timeline. Not for surveillance. For understanding.

When a teacher opens a student's WorkHub in Koan, they do not see a finished product and a grade. They see the path. The three drafts. The moment the student abandoned a weak thesis. The five-minute pause before a breakthrough. The question from Aidan that prompted a new direction. The learning is not inferred from the output. It is visible in the record.

Building for the Journey, Not the Destination

The lesson from the Canvas episode is not that agentic AI is dangerous (though it can be) or that educational AI is helpful (though it can be). The lesson is that any system built entirely around deliverables will eventually be gamed by machines. This was true of assembly lines. It is true of software. It is now true of education.

The OECD's Digital Education Outlook 2026 documented this pattern with precision: students using general-purpose AI tools show a 48% performance boost that collapses into a 17% deficit once the tool is removed. The tool carried them. It did not teach them. But purpose-built Socratic tools, designed around questioning rather than answering, produced sustained gains that persisted after the tool was gone.

The difference is not the power of the model. It is the philosophy of the interaction. Tools that generate outputs create dependency. Tools that make thinking visible create capacity.

A Question Worth Asking

Seven in ten middle and high school students now say they worry that AI is eroding their critical thinking skills. That statistic, from a new RAND Corporation survey, should stop us in our tracks. The students themselves can feel it. They know something is being lost. They are asking for help, in the only way a survey allows.

So here is the question this moment demands, not just of Canvas or Instructure, but of everyone building technology for classrooms:

If a machine can complete your course without learning anything, what does that tell you about what your course was measuring in the first place?