Harvard's agentic-Ai paper — and the student side of the apprenticeship gap

Harvard Kennedy School's Mossavar-Rahmani Center published a paper this month on the future of work in the age of automation, augmentation, and agentic Ai. The full report is here: https://www.hks.harvard.edu/centers/mrcbg/publications/future-work-age-automation-augmentation-and-agentic-ai.

The employer-side headline is the one that's been quoted everywhere: entry-level cognitive work is being automated first, and the labor-market data is already showing it — meaningful declines in openings and hours for the roles that used to be the on-ramp into a career. The uncomfortable half of the argument, though, is on the student side of the pipeline: if AI does the first three years of the job, no one does the first three years of the learning. The apprenticeship loop breaks from both ends.

The apprenticeship loop, in one paragraph

Expertise doesn't come from consuming correct answers. It comes from attempting, being wrong, being corrected, and trying again on a slightly harder version of the same problem. That's true for a junior analyst writing their first memo, and it's true for a 16-year-old working through a chemistry set. Take out the attempt and the correction, and what's left is a track record of correct-looking output with no underlying skill. Harvard's paper describes this happening at the employer scale. It's already happening at the student scale, one homework tab at a time.

The Brown data point, one layer up

“

40 perfect take-home midterms. A 48-point average on the in-person final. That is the apprenticeship loop breaking in real time — three years earlier than Harvard is describing it.

Late last year a Brown professor published a class where 40 of 86 students scored a perfect 100 on a take-home midterm and then averaged 48 on the in-person final. Same students, same material, three weeks apart. That's the student-scale version of the paper's warning: AI removed the practice reps, everyone looked competent, and the gap didn't show up until the moment there was no Ai in the room. The paper is describing this happening across an economy. Brown is what it looks like inside a single classroom.

What an honest Ai study tool has to do

If the pipeline is collapsing from both ends, a study tool can't just be a smarter answer box. It has to protect the three steps that build skill: the attempt, the honest signal about how sure the answer actually is, and the second-pass check on a different day. LemonSugar Ai is built around exactly those three:

• Practice — the moment a student gets an answer, the next tap is a 3-minute recall drill on the same concept with no Ai in the loop. That's the attempt the take-home midterm skipped.
• Proof Receipt — every answer ships with a chip showing where it came from and how sure the model actually is. Low-confidence stays labeled low-confidence. Students learn calibrated trust, not blind trust.
• Memory — the notes you upload, the weak spots the quizzes surface, and the topics you keep coming back to compound over time. Sugar's brain is the student's, not the vendor's.

Same paradox, other side of the desk

Harvard's paper is aimed at the people running companies: don't automate away the on-ramp you'll need in five years. LemonSugar Ai is aimed at the people about to step onto it: don't let Ai run the reps for you, because when the room goes quiet — the final, the interview, the first day of the job — the reps are all you have. Route the query to the cheapest capable model. Show the confidence honestly. Put the recall drill one tap away. That's the on-ramp we're building.