The apprenticeship AI does not replace
Shopify is expanding internships while embracing coding agents. The more useful workforce question is how quickly a new engineer learns to own the result.
Editorial image for The apprenticeship AI does not replaceVIBECIO Editorial Desk ·
The argument for cutting entry-level engineering roles often starts with a neat substitution: agents can produce routine code, junior engineers used to produce routine code, therefore the junior role is disappearing. That skips the reason an organization develops engineers in the first place. Someone has to acquire the knowledge to diagnose the next unfamiliar failure, challenge a design assumption, and eventually take responsibility for a system. Routine work was one route into that knowledge. The route can change without the destination becoming unnecessary.
Shopify is testing a different workforce bet. In his June 1 essay, engineering leader Farhan Thawar says the company is hiring 1,000 interns again in 2026. He describes River, an internal coding agent used in public Slack channels, as a way to expose newcomers to real engineering decisions. Shopify reports River’s pull-request merge rate rising from 36% to 77% over two months as employees improved its context. That is a company-reported operating measure, not a controlled estimate of developer productivity.
The output is not the apprenticeship
A high merge rate can reveal that an agent’s proposals are becoming more usable. It cannot, on its own, show that an intern understands a permission boundary or can investigate an incident. Those are different outcomes. A program that rewards only accepted patches may train participants to become excellent at producing the artifact while leaving their understanding of the system uncertain.
A stronger apprenticeship makes that understanding observable. Before changing a service, ask the engineer to trace a request through its dependencies and identify which system owns the final state. After the change, ask what would fail if an assumption stopped being true. Then introduce a controlled variation: an expired credential, a delayed event, or a duplicate request. The lesson is whether the person can interpret the result and choose the next action.
This need not become an oral examination attached to every commit. It can be part of a weekly working session around one meaningful change. The mentor supplies context and challenges the explanation. The apprentice uses tools to investigate, proposes a fix, and demonstrates the evidence. Repeating that cycle on progressively less familiar work creates a record of growing responsibility.
Measure the whole path to useful work
Microsoft’s EngThrive framework distinguishes speed, ease, and quality, with sustainable work as a further constraint. Tim Bozarth and David Liu argue for studying the engineering system rather than rating individuals by activity counts. Their framework offers a useful direction for workforce decisions: observe where work slows down and where effort is displaced, not merely where code appears faster.
For an onboarding cohort, track time to the first useful change alongside reviewer effort and the amount of rework after acceptance. Later, examine how often an engineer can resolve a bounded production issue without a mentor taking over. Include the quality of escalation: recognizing that a task exceeds one’s authority or understanding is part of competence. These measures describe a learning system rather than creating a leaderboard of token use.
The productivity evidence keeps moving
METR’s July 2025 randomized study found that 16 experienced open-source developers took 19% longer on assigned tasks when allowed to use the early-2025 AI tools they selected. The study covered 246 tasks in familiar repositories. Its authors explicitly warned against generalizing that result to most development work or future systems. It did not study whether an internship program should expand.
The February 2026 follow-up makes a simple anti-AI reading even less tenable. METR reported that developers increasingly declined tasks or participation that required working without AI, complicating the sample. Concurrent agent use also made time accounting harder. The researchers considered the new data an unreliable measure of the current productivity effect and planned a different design. A workforce strategy needs to accommodate that uncertainty rather than borrow a convenient percentage.
Make responsibility grow with capability
Shopify’s public account keeps human ownership explicit: the person submitting a pull request owns the code, and a human reviews it. It also warns that tool access does not replace mentorship. The practical question for other employers is what infrastructure would make such ownership teachable in their own environment.
Give newcomers bounded work that matters, access to the reasoning behind decisions, and a safe place to see how systems fail. Give mentors time to review explanations as well as outputs. Expand production authority when the engineer can demonstrate understanding and recover from foreseeable errors. Faster implementation can increase the number of learning cycles available; it can also let someone skip them. The apprenticeship has to be designed to capture the first effect without quietly accepting the second.
Sources