Microsoft CEO Satya Nadella tells Gergely Orosz that cheap AI code makes verification the hard part of software engineering, and that every company should own its own AI stack. The hour-long conversation, recorded in Redmond for The Pragmatic Engineer, covers how Microsoft lost the first round of agentic coding to Cursor and Claude Code, what a “software factory” looks like inside the company, the voluntary exit offers, the junior hiring slowdown, open and local models, and why Nadella thinks agents have to be treated as insider risk.
TLDW
Nadella argues that code has gone from being the input of software engineering to its output, so specs, intent, tests and verification loops are now where engineers earn their keep. He admits Microsoft was slow when Anthropic’s models made the command line the main form factor for coding, and says AI product teams cannot afford to be in love with what they built. On the business side, he says every firm needs its own harness, context layer and private evals so it can swap any model, and a “trust architecture” that keeps authority away from unreliable intelligence. He frames the voluntary exit offers and the slowdown in junior hiring as a reconfiguration of teams rather than a long-term trend, and closes with one piece of advice: do more, talk less, and get back into school mode.
Thoughts
Nadella’s acid test for a company in the AI era is simple: you should be able to replace any model you use and still improve against your own private eval. If you can’t, he asks, “how are you a sovereign firm?” This is the most useful framing in the interview because it turns a vague worry about vendor lock-in into an architecture rule. Keep the harness, the context and memory, and the action space (plugins, skills, connectors) separate from any single model, and keep the evals that define success for your business in-house. His analogy is the relational database. A bank could always migrate its data off a database vendor that stopped shipping updates. If your firm’s learning lives in the weights or embeddings of someone else’s model, he asks, how do you migrate that out? Orosz admits he had assumed companies building their own harnesses and inference stacks were mostly engineers enjoying new toys. Nadella’s answer is that they are doing their companies a strategic service. It is worth noticing that this advice also happens to sell Microsoft Foundry, which hosts the Copilot harness and many models as a managed service. That does not make it wrong. A CEO who partners with OpenAI and trains his own MAI models has every reason to want customers who can switch models freely.
The second idea is the sharper one, and it comes right after. Nadella says scaling laws bring new capability but not alignment, and that we have no mechanistic understanding yet, so models are getting “more and more capable and less and less reliable”, which he calls the most dangerous place to be. His conclusion is to separate authority from intelligence architecturally and to treat agents as insider risk. Chain of thought should be in human language, auditable and logged so it can’t be tampered with, because reward hacking can happen at test time, not just in training. This is a striking thing for the CEO of one of the largest AI vendors to say on the record. It also gives engineers a concrete design principle that does not depend on trusting any model: an agent’s permissions should never exceed what you would grant a new contractor you cannot fully vet.
On junior hiring, Nadella offers a theory that is more interesting than “AI replaced entry-level jobs.” He says an AI-native company starts with a token budget and a human, adds the next human only when the first one maxes out their tokens, and gives each new person and their factory a budget. Incumbents are now trying to rebuild their teams in that shape, and he thinks the hiring slowdown may be the pause while those factories get built so new people can plug in with high leverage. It is a hopeful reading, and he is careful to say it is too early to draw long-term lines. It sits next to some hard facts. Microsoft announced its Voluntary Exit Offer in late October, and Business Insider reported that around 60,000 employees in the US, Canada, the UK and India are eligible. Nadella says the offer is about giving people agency to move to where their long-term economic opportunity is, and admits “we have created enormous anxiety” with talk of a singularity and no jobs. The honesty is welcome. The harder question, which the interview doesn’t fully answer, is how many of those reconfigured teams will have slots for new graduates at all.
Underneath everything is the line Orosz used to open the episode: when code generation is cheap, verification is hard. Nadella describes building tests from every failure, using a different model family to critique the first so one family’s blind spots don’t become yours, and feeding production traces back so an agent can propose a fix, verify it and deploy it through the full test suite. He jokingly calls that last loop his version of recursive self-improvement. The practical message for working engineers is that the skills haven’t been devalued, they’ve moved. Design docs, clear intent, test harnesses, CI/CD and observability used to be the chores around coding. Now they are the work, and on brownfield systems like Windows the build system and CI are the hardest part of all.
Key Takeaways
- Code has moved from input to output. Specifications, intent and design docs are now the main things engineers write and review.
- With cheap code generation, verification is the hard part: turn every failure into a test case, use a separate model family as critic, and confirm a task is truly done.
- Nadella admits Microsoft did not have an immediate answer when Anthropic’s models made the CLI the dominant coding form factor and IDEs faded within about six months.
- AI-first product building is “unforgiving.” When model jumps happen nonlinearly the form factor has to change quickly, so teams should not be in love with what they built.
- He expects a possible rebirth of the IDE, because decision engines can be embedded in software like intelligent functions in a more token-efficient way.
- Decision engines are a new primitive, the “stored procedure of the intelligence era,” meant to become as reliable and boring as relational databases.
- Production traces feeding an automated find, fix, verify and deploy loop is what Nadella calls his own version of recursive self-improvement.
- Microsoft’s Azure networking team built an agentic system to run operations across roughly 500 fiber operators, connected directly to its code repos.
- Brownfield work is the real world. Coding agents have pushed GitHub’s scale so high that Microsoft is reworking the git protocol and storage system.
- The voluntary exit offers are framed as giving employees agency to move to where their economic opportunity is, as teams change size and shape.
- Forward deployed engineer is one of the fastest growing jobs at Microsoft, and Nadella wants sales, support and FDE roles trained alongside software engineers.
- The junior hiring slowdown may be temporary while companies rebuild teams around a “token budget and a human” model.
- Every firm should own its harness, context, memory and action space separately from models, and should be able to swap any model while improving on a private eval.
- Agents should be treated as insider risk, with authority separated from intelligence and chain of thought kept auditable and logged.
- Hybrid intelligence means routing some sub-agents to local models on the PC, which Nadella calls “unmetered intelligence,” while cloud demand keeps growing.
Chapters
02:12 From Open Source and VS Code to Claude Code
Nadella says Microsoft’s open-source commitment predates his time as CEO, and that VS Code taught the company what building in the open really means, with product managers who listened to the community rather than to a vice president. Orosz asks how a company with developer DNA missed the 2025 shift to Claude Code. Nadella says Copilot’s early code completions reaffirmed his belief in scaling laws, Cursor did a good job on agent mode, and Anthropic’s models then made the CLI the form factor. The lesson he draws is to be ready to rewire everything when the next model arrives.
09:52 Decision Engines as a New Primitive
Nadella compares decision engines to macros in Excel: intelligence embedded as a primitive inside code. His example is an Xbox system for triaging community feedback, where you want a decision rather than freshly generated code each time. He loves the goal of making these engines as boring and reliable as relational databases, and agrees with Orosz that they could become the stored procedures of the AI era.
12:32 Verification and the Software Factory
Inside Microsoft, the GitHub Copilot runtime team had a single developer maintain and migrate a whole runtime, and another team runs a software factory with roughly a dozen engineers. Nadella says he would love to join one of these teams, and praises the preceptorship program Mark Russinovich and Scott Hanselman are building for early-career engineers. New hires find codebases easier to learn because everything in a software factory is documented. He describes verification loops, critique from a separate model family, and the Azure networking team’s agentic operations system.
20:45 Bringing AI to Brownfield Products
Orosz asks how product teams with real customers can adopt any of this. Nadella says there is no permission to do things that don’t add customer value, and admits Microsoft built “random AI things” that were just cool. AI models now find vulnerabilities in Windows, so a new pipeline runs from crash traces to fixes on top of the existing build system. He argues this differs from adopting React because the runtime, framework, IDE and stack are all changing at once, and a new model could change the regime again at any moment.
26:29 The Changing Role of Software Engineers
Nadella says the same skills are still required, but the order of operations and the toolchain have changed, so expertise needs new leverage rather than being devalued. He recalls meeting the first class of college graduates who never knew an IDE without GitHub Copilot. Experienced people with depth and curiosity, he says, are flourishing.
29:25 Voluntary Exit Offers and AI Anxiety
Asked directly whether the voluntary offers are a prelude to layoffs, Nadella says the company wants all 200,000 people to flourish, but they will be configured differently, with new team sizes and workflows. The offer gives people agency, including the option to vote with their feet if they are on a team that isn’t transforming. He acknowledges the industry has created enormous anxiety and points to forward deployed engineering as one of Microsoft’s fastest growing jobs.
36:35 Junior Hiring
Nadella says the first job is to reorganize existing staff into the right places, and only then add industry and new college hires. He thinks the slowdown may last only as long as it takes to build the new software factories, comparing incumbents to AI-native companies that grow one human and one token budget at a time.
42:36 Open Models and Owning Your AI Infrastructure
Drawing on Hayek’s idea that knowledge is distributed throughout an economy, Nadella argues there should be as many frontier models as there are firms, with each firm holding both human capital and token capital. Companies can start on a frontier model, collect traces, then fine-tune a smaller model to cut costs. He names the “reverse information paradox” as the risk to avoid: don’t leak your value, and stay in control of your own learning loop.
51:21 A Trust Architecture for AI Agents
Because models are gaining capability faster than reliability, Nadella says firms must separate authority from intelligence, monitor agents as insider risk, and keep chain of thought auditable in human language. He previews Microsoft’s Ignite conference as the place to show that there are workable blueprints for all of this.
53:48 Local Models, MAI and the Hybrid Intelligence Era
Nadella says last year’s frontier models already run on a laptop. A model router can send some sub-agents in a swarm to local models while others use the cloud, which he calls unmetered intelligence. Microsoft will keep using OpenAI models through its partnership while building its own MAI model family, aiming for new points on the frontier that are efficient in both training and deployment. He closes with advice for engineers: change your ratio of talking to doing, and get back to school mode.
Notable Quotes
“In a world where code generation has become cheap, what has become hard is verification.”
Satya Nadella, on the software factory
“When model jumps happen nonlinearly, the form factor has to change quickly.”
Satya Nadella, on losing the first round of agentic coding
“We have created enormous anxiety, enormous anxiety in people sort of feeling that they don’t have control of their own destiny.”
Satya Nadella, on how the industry talks about AI and jobs
“What is an AI native company? They start with a token budget and a human, right?”
Satya Nadella, on the junior hiring slowdown
“You should be able to replace any model and still hill climb on an eval that’s your private eval.”
Satya Nadella, on his test for a sovereign firm
“If your intelligence has authority and it’s unreliable, again, you’re taken, right? You have a massive insider risk problem.”
Satya Nadella, on trust architecture for AI agents
“When am I at my best? It’s when I’m in the school mode.”
Satya Nadella, on advice for engineers
Watch the full conversation here.
Related Reading
- Microsoft is dogfooding AI dev tools’ future Gergely Orosz’s deep dive on how Microsoft teams build with AI tools.
- What are Forward Deployed Engineers? background on the role Nadella calls one of Microsoft’s fastest growing.
- The Use of Knowledge in Society (Wikipedia) the Hayek essay behind Nadella’s case for as many models as firms.
- sn scratchpad Satya Nadella’s own blog.
- Hit Refresh by Satya Nadella, his account of the culture change that led to Microsoft’s open-source turn.