Microsoft's PRAISE Program: A New Paradigm for Training Software Engineers in the AI Era

When AI can complete programming assignments that once took hours in mere seconds, how should software engineering education respond? Microsoft and Carnegie Mellon University are exploring the answer with the "PRAISE" program — short for "Preceptorship for AI-accelerated Software Engineering" — which seeks to redefine the training path for newcomers in the AI era through a "preceptorship" model rather than traditional internships.

The Educational Dilemma Under AI's Impact: When Assignments Can Be "One-Click Completed"

In an interview on Microsoft's Command Line blog, Dr. Michael Hilton, a teaching professor at Carnegie Mellon University, shared a key turning point: "When AI can complete all of our assignments in one shot, academia reaches a real inflection point."

This moment arrived roughly between the fall 2025 and spring 2026 semesters. Professor Hilton teaches upper-level software engineering courses, and he observed that students no longer needed to think deeply — they could simply copy and paste to finish their assignments. This exposed a long-standing weakness in the educational system: teachers had relied on the assumption that students could either do the work themselves or not do it at all, but now a third option has emerged: letting AI do it.

The experience of Julia Gao, a Microsoft software engineer, confirms this shift. During her internship in 2024, AI was mainly used for basic tasks like syntax conversion; but after going full-time, "the introduction of models like Cloud Opus completely changed the way we work — we now use multi-agent workflows to handle all coding." Her role shifted from "the person writing code" to "the person reviewing code."

Context Window refers to the maximum number of tokens a large language model can process in a single inference, including both the input prompt and the model's output. It determines how much information the AI can "see" when answering questions or generating code. The larger the context window, the more code, documentation, and conversation history the model can reference simultaneously, and the more likely it is to make judgments consistent with the overall context. However, a context window is not the same as understanding the system: even with a sufficiently large window, the model still lacks grasp of the runtime environment, deployment architecture, historical decisions, and implicit constraints. This is precisely why Hanselman in the interview distinguishes between "AI's context window" and "the larger context in which the solution exists" — the former is a technical parameter, the latter is engineering reality.

PRAISE's Core Innovation: Redefining the "Mentor" Role

Scott Hanselman emphasized a key idea in the interview: "The intern label says 'you're here to learn,' while PRAISE gives senior engineers the label of 'Preceptor,' making them realize their job responsibility has changed — not just delivering code, but 'delivering senior engineers.'"

Professor Hilton analyzed the importance of naming from an academic perspective. "As a scholar, I'm qualified to be pedantic," he laughed, "but what really matters isn't the word 'preceptor' itself, but that there's a clear vision and plan behind it. If you just pick a word randomly from the dictionary, it's meaningless."

Julia shared her personal experience: "PRAISE is much better than ordinary mentorship. The weekly structured AI lab is the starting point for discussion, and I can ask my mentor anytime — not just technical questions, but about ways of thinking. He'll start from the project's initial steps and tell me how he would think about it."

Preceptorship originates from medical and legal education traditions, referring to long-term, structured guidance provided by senior practitioners to newcomers on a one-on-one or small-group basis, aimed at cultivating judgment — distinct from traditional internships that focus on task distribution. In medicine, attending physicians mentoring residents not only teach procedures but also convey the thought process behind clinical decisions. PRAISE borrows this concept to make senior software engineers aware that their role is shifting from "code deliverer" to "engineer cultivator." Compared to traditional mentorship, preceptorship emphasizes a clear curriculum design, periodic feedback, and demonstration of thinking, rather than a loose "come ask if you have questions" arrangement.

In an Era of Depreciating Coding Skills, What Is the Core Competency?

The most insightful discussion in the interview revolved around "context." Hanselman pointed out the multiple meanings of the word: AI's context window is completely different from the larger context in which a solution exists.

Julia explained from an engineering practice perspective: "AI is very good at solving a single problem in a single codebase, but it's not good at situating itself within the broader context of the entire system. It might fix one problem while introducing new ones elsewhere."

Professor Hilton introduced the concept of "replacement value": "If AI can write code based on a task description, and that's all you would do, then you're in trouble. But a generic prompt usually doesn't include context — are we sending network packets to a device across the street, or to the moon? That context determines the requirements for verification, performance, and security."

Multi-Agent Workflow refers to an architecture in which multiple AI agents collaborate on complex tasks with divided roles. Each agent can take on different roles such as planning, coding, reviewing, or testing, coordinating through message passing or shared state. This pattern allows AI to handle longer task chains than a single prompt, but it also brings new challenges: handoffs between agents may lose context, errors can amplify along the chain, and the final output still requires verification and integration by human engineers. When Julia Gao mentions "using multi-agent workflows to handle all coding," it shows that engineers' focus has shifted from writing code themselves to orchestrating, reviewing, and being accountable for results.

From Writing Code to Reading Code: A Fundamental Shift in the Educational Paradigm

In an op-ed published in The New York Times titled "You're a Computer Scientist. Don't Panic," Professor Hilton and Dr. Mary Shaw put forward a core argument: when typing is no longer the bottleneck in software engineering, reading, understanding, and critical thinking become the key skills.

He compared software engineering to civil engineering: "Civil engineers don't need to personally mix concrete and pour bridges, but that doesn't mean they lack knowledge or skills. As the cost of writing code approaches zero, the engineering cost of ensuring problems are correctly solved has not decreased."

Professor Hilton particularly emphasized: "Reading and critiquing code is itself an independent skill, different from writing code. History has led us to assume people will naturally acquire it while writing code, but in the AI era, we need to teach this skill directly. Just as my son reads The Odyssey and writes a review but isn't asked to write an epic — computer science education also needs a 'code appreciation course.'"

Code Review and PR Review are core practices for ensuring quality in software engineering. A PR (Pull Request) is the mechanism by which a developer requests that code changes be merged into the main branch; reviewers must evaluate logical correctness, maintainability, security, and style consistency. With the proliferation of AI-generated code, reviewers often face unfamiliar and seemingly reasonable changes, and must rely on their understanding of system context to judge their real impact. AI can also assist with PR reviews by suggesting improvements, but the suggestions themselves may expand the scope of changes, requiring human judgment on whether to adopt them. Therefore, review capability has risen from an "incidental skill" to an independent and critical core competency.

Bottleneck Shift and Future Outlook

Julia described the changing bottlenecks in daily work: "Reading code, understanding AI's changes, verifying whether they meet expectations, and waiting for builds to complete. AI is now also used for PR reviews, and sometimes its suggestions expand the scope of work — I need to judge whether to adopt them. The actual writing of code is no longer time-consuming; getting code ready for production is what matters — because you're responsible for it."

Professor Hilton remains optimistic: "We shouldn't be doomsayers. More people will use software to solve more problems. The key is to directly teach engineering judgment and taste — things we used to assume would be naturally acquired over years of coding."

Hanselman concluded: "A senior engineer's job is to deliver products, while a mentor's job is to deliver senior engineers." This may be precisely the most core value proposition of the PRAISE program.

One-line Summary

Microsoft and CMU's PRAISE program explores software engineering education in the AI era, focusing on cultivating judgment rather than coding speed.

Paragraph Summary

Facing the impact of AI completing programming assignments in seconds, Microsoft and Carnegie Mellon University have partnered to launch the PRAISE program, replacing traditional internships with a "preceptorship" model that shifts senior engineers from delivering code to delivering senior engineers. The article points out that the educational system has long assumed students could either do their assignments themselves or not at all, but AI provides a third option, exposing teaching weaknesses. The coding bottleneck has shifted from writing code to reading code, understanding system context, and verifying AI changes, with engineers' core value becoming engineering judgment and taste. Professor Hilton advocates teaching reading and critiquing code as independent skills directly, analogous to civil engineers not needing to pour concrete themselves. PRAISE uses weekly structured AI labs and demonstrations of thinking to help newcomers build system-level understanding and accountability in an AI-accelerated environment.