Satya Nadella: When Code Is Cheap, Verification Becomes the Work

Microsoft’s CEO says teams need stronger verification and control over their AI systems as coding agents change how software is built.

SNSatya NadellaInvitado · Microsoft
Canal
The Pragmatic Engineer
Publicado
16 nov 2026
Vídeo
1:03:39
Lectura
10 min
Original
inglés
Abrir en YouTube
Al desplazarte, el vídeo se reduce a la esquina inferior derecha y sigue reproduciéndose. Los subtítulos siguen tu idioma de lectura.
Idioma de lecturaDisponible en 2 idiomas
Este artículo aún no está en español, así que se muestra el original en inglés.
RESUMEN

En 30 segundos

As AI makes generating code cheaper, specifying intent and verifying results become harder, says Microsoft CEO Satya Nadella. He describes teams turning failures into tests, using separate models as critics and feeding production traces back into fixes. He also argues that companies should keep their harness, context, models and permissions separable so they can change providers and protect their knowledge. On hiring, he says a slowdown in junior recruitment may reflect teams reorganizing, while acknowledging anxiety about changing roles.

  1. 1Model jumps can change the product form factorNadella says the move from completion to agent mode and then command-line coding shows why AI product teams must be ready to rebuild around new model capabilities.
  2. 2Cheap code makes verification the hard partHe points to turning failures into test cases and checking that generated fixes really work, rather than treating a large volume of generated code as success.
  3. 3Specifications and documented context become inputs to the software factoryNadella says teams need clear intent, small verifiable tasks and critics. Well-documented systems can also make it easier for new engineers to learn a codebase.
  4. 4Existing products need a staged transition, not a clean-slate factoryHe describes Windows vulnerability fixes and GitHub’s scaling work as changes to existing build, storage and delivery systems that require investment and a careful order of operations.
  5. 5The junior hiring slowdown may reflect a period of reorganizationNadella says teams are being reconfigured around new workflows and cautions against treating one year of hiring as a settled long-term trend.
  6. 6Keep context, harness and models independent enough to replaceHe argues that firms should own their learning loop and use private evaluations to test whether a different model can deliver the outcomes they need.
  7. 7Separate an agent’s intelligence from its authorityNadella proposes monitoring agents as an insider risk, with runtime safeguards and auditable records, because greater capability does not ensure reliability.

Conclusiones

  1. Treat every failure as a potential test and verify agent-generated changes before deployment.
  2. Keep model choice, context and memory, agent harness and action permissions separable.
  3. Evaluate models against private measures of the outcomes your organization needs.
  4. Separate an agent’s intelligence from its authority to act, Nadella says.
  5. Bring production traces into a loop that can reproduce, fix and verify failures.
  6. Nadella urges engineers to learn the new workflows hands-on, including in teams already using them.
TEXTO COMPLETO

Lee el texto editado por capítulos

Solo la transcripción
01

Open source and learning to build with a community

Nadella describes Microsoft’s open-source shift and VS Code as a lesson in incorporating community feedback.

Gergely Orosz recalls working at Microsoft when Satya Nadella became CEO in 2014. He asks why a company associated with closed source embraced projects such as TypeScript and VS Code, acquired GitHub and moved to Git.

Nadella says Microsoft had contributed to Linux before he became CEO, but wanted to make its commitment unmistakable: bring workloads such as SQL Server to Linux, contribute to the ecosystem and earn the open-source community’s trust. He argues that perceptions of companies’ contributions can diverge from the underlying facts.

Orosz says the free and open VS Code made a particularly large impression on developers. Nadella says its community-centered development was intentional: the team learned to use daily feedback from users, not only direction from management, to build the product.

02

Why model advances change developer tools

The discussion moves from autocomplete to agent mode and command-line coding. Nadella also expects decision engines to become reusable software primitives.

Orosz notes that GitHub Copilot was early to AI coding, while Cursor later built on VS Code and Claude Code brought a command-line workflow. He asks why Microsoft did not immediately respond when the dominant interface shifted.

Nadella recalls that early code completion was convincing despite skepticism. His lesson is that when model capability jumps nonlinearly, an AI product’s form factor may have to change quickly. A team should build for current capabilities without becoming attached to an interface that the next model may make obsolete.

He says Cursor did well with agent mode, which allowed developers to assign small tasks rather than accept individual completions, and Microsoft subsequently caught up. Anthropic’s model then helped make command-line coding and fuller task handling viable; the IDE became less central for some workflows.

The next interface, in his view, may emphasize specifications, Markdown and other inputs to a software factory. Yet he also sees a possible renewed role for an IDE if small decision engines can be embedded efficiently in software. Different model capabilities can support several form factors.

Orosz asks about decision models and the limits of language models on reliable programmatic tasks. Nadella sees decision engines as a new primitive, comparable to what macros added to spreadsheets: software can make a bounded decision without repeatedly generating a whole application.

He cites an Xbox system for managing community feedback as an example of an application that needs decisions about incoming material. An embedded component could make those decisions more efficiently than generating code again for each use.

Nadella compares the aspiration to a relational database: a dependable primitive made many applications possible. Orosz suggests the analogy of a stored procedure for an era of intelligence. Both see room for new building blocks beyond general-purpose code generation.

03

Verification and the emerging software factory

Generating code shifts work toward specifications, critique, tests and feedback from production. Nadella wants to learn these practices from teams using them.

In a world where generating code has become cheap, Nadella says, verification has become difficult. He wants to learn from teams that increasingly produce code as an output: the process still needs a particular order, not uncontrolled generation.

He describes converting each failure into a test case and checking that the test genuinely verifies the fix. He notes that a Rust migration had to be reverted, illustrating that faster generation brings disadvantages as well as benefits.

Nadella says he would like to join a hands-on engineering team again. He is interested in a Microsoft preceptorship program because early-career people need a way to join such teams, climb the learning curve and see how experienced engineers work.

Being close to customers and building in the open also remain ways to learn. Orosz observes that technical leaders sometimes step down into smaller teams to learn new tools firsthand; Nadella agrees that hands-on work matters during rapid change.

Asked to define a software factory, Nadella says he is still forming his understanding. He points to a GitHub example of an individual maintaining a runtime and to a small product team that has written about its development process.

If code is now more often an output, specification and intent become especially important inputs. Design documents and team alignment were already valuable; the new workflow makes their role more visible.

A new engineer told Nadella that entering a codebase can be easier when the factory requires documentation. An agent can answer questions patiently, helping the newcomer recover information that otherwise might exist only as tribal knowledge.

For agentic coding, he advocates small tasks, checks that the work is really finished and a critique step. He is interested in a critic from a different model family, which might catch problems that the first family’s tendencies could propagate.

This remains software engineering, he argues, although the level of interaction and team organization change. He compares the shift to moving from hand-written assembly to compilers: engineers express intent at a higher level of abstraction.

Another loop brings production traces back into development. A team can use logs to understand a failure, create and verify a fix, run tests and deploy it. Nadella sees that end-to-end verification loop spreading across Microsoft teams.

He describes an Azure networking team managing relationships with hundreds of fiber operators. Work that previously involved emails and ad hoc processes is being connected to repositories through an agentic operations system. The technical team builds and manages the combined operations and software system.

04

Changing existing products and engineering roles

Windows and GitHub must adapt within established systems. Nadella says expertise remains useful even as workflows and abstraction levels change.

Orosz asks how teams with real customers and established products can adopt these methods without abandoning their obligations to build a new factory. Nadella says work must add customer value and admits Microsoft has made mistakes by doing AI projects because they seemed exciting.

He points to Windows vulnerability and crash traces: AI can discover more vulnerabilities, so the team needs a pipeline for fixes in an existing product. The Windows build system remains the Windows build system; that makes the task a brownfield transition, not a clean-slate project.

GitHub faces a different scaling challenge as coding agents generate much more activity. Nadella says Microsoft is working on the Git protocol, storage and surrounding infrastructure. Existing codebases require gradual hardening of source control, builds and delivery, and the order of changes matters.

Orosz compares adoption to introducing React gradually when teams added features. Nadella sees a larger shift: not just a new framework, but potentially the runtime, interface, stack and order of operations. Another model breakthrough could change that configuration again.

Orosz asks how engineers should reconcile the loss of work they spent years mastering. Nadella argues that the same depth and expertise remain useful, but engineers must exercise them through different tools and workflows rather than assuming the expertise has lost value.

He has met graduates who entered college with GitHub Copilot already available and are native to that toolchain. He also says experienced Microsoft engineers who combine domain depth with curiosity about new workflows are flourishing, and that the company should give people ways to build marketable skills.

05

Voluntary exits, anxiety and early-career hiring

Nadella frames exit offers as giving employees agency amid restructuring. He expects roles to change and says it is too early to infer a lasting hiring trend.

Orosz asks directly about Microsoft’s voluntary exit offers and whether they precede AI-related layoffs. Nadella says future opportunities and team configurations will differ, and frames the voluntary program as giving employees more agency over their long-term economic opportunities.

Nadella says Microsoft wants employees to find opportunities internally, including when they feel stuck on a team that is not transforming. He cannot promise that work will stay the same, but says the company should create opportunities and help people grow through the change.

Orosz interprets the offer as a way to leave rather than follow a period of intensive change. Nadella agrees while acknowledging that the way the industry has discussed AI has created substantial anxiety about people’s control over their futures.

He says forward-deployed engineering is among the fastest-growing roles at Microsoft. Account, customer-success, technical-support and engineering teams may need related hands-on learning; he cites LinkedIn’s “full-stack builders” as an example of combining roles.

Orosz asks whether Microsoft is still investing in new graduates when many companies have slowed junior hiring. Nadella says the company first needs to reconfigure existing people and teams, after which it can add experienced hires or graduates to the new arrangement.

He suggests the hiring slowdown may partly reflect time spent building those workflows. An AI-native team, he says, might start with a person and a token budget, add people as it expands, and reorganize around those resources. He cautions that it is too early to project a single year into a permanent trend.

06

Open models and control of a company’s learning

Nadella argues that organizations need model choice, separate context and private evaluations. He describes Microsoft Foundry as one way to combine those parts.

Orosz asks whether open models will follow a trajectory like open source as companies deploy their own inference. Nadella expects both closed competition and open alternatives, with open weights offering flexibility.

Rejecting the premise that one frontier model must contain all useful knowledge, Nadella invokes the idea that knowledge is distributed throughout the economy. He imagines firms combining human and token capital, with their own knowledge in open weights, context or systems using multiple closed models.

A company can retain usage traces, train a smaller model and combine it with other models. Nadella describes Copilot as a multimodel system and imagines a decision layer choosing a model while context, memory and available actions remain separate from any single provider.

If the harness, model and memory are tightly coupled, he asks, how does an organization leave the provider or know what it owns? He argues that banks, healthcare firms and other organizations need an architectural approach that preserves control over their learning and differentiating knowledge.

Orosz observes that many companies already build their own harnesses, context layers and inference systems. Nadella agrees that keeping replaceable components can be strategic, not merely an engineering experiment.

He describes Microsoft Foundry as offering a managed harness while also allowing open-source alternatives. Customers can connect multiple model endpoints, retain their own context and memory, and keep tools, skills and connectors separate from a particular model.

Private evaluations matter more than public benchmarks for an organization’s own outcomes, Nadella says. A customer might begin with a frontier model, collect traces and then tune a smaller model for a more predictable workflow to reduce cost. He presents Foundry as supporting that kind of iteration.

His analogy is a relational database: a company can buy software while keeping its own data and business logic, and can migrate if a vendor stops updating a product. He asks how a company would move its memory if that information were embedded in only one provider’s model.

Nadella’s practical test is whether an organization can replace a model and still improve against its private evaluation. He says the harness, context and model should be distinct enough to preserve control over the firm’s learning loop.

07

Trust architecture and hybrid intelligence

Agents need limited authority and monitoring. Nadella sees local and cloud models working together through routing.

A second enterprise decision is trust. More capable models are not automatically more reliable or aligned, he argues, so intelligence must be separated from authority. An agent with both broad permissions and unreliable behavior is an insider-risk problem.

He calls for runtime safeguards, monitoring, observability and auditable records of reasoning. He also worries that systems can pursue the wrong reward during inference, not only during training, which makes controls at execution time important.

Orosz hears a large amount of new engineering work in these requirements. Nadella agrees and says companies need to connect architecture choices to enterprise value rather than focusing only on novelty.

Orosz asks where local models belong when cloud inference and expensive hardware dominate. Nadella says distributed computing should use capability both in the cloud and on devices. He cites progress in chips and systems that allows more capable models to run on laptops and desktops.

He describes routing among local and cloud models: a swarm of subagents could use local models for some tasks and cloud models for others. He calls this hybrid intelligence and describes local capacity as a form of unmetered intelligence.

08

Microsoft’s model strategy and advice for engineers

Microsoft plans to use multiple model families alongside its own work. Nadella recommends hands-on learning over following the hype cycle.

Orosz notes that Microsoft’s MAI coding model can run locally even though it is closed. Nadella says taking coding models to laptops is a priority, while Microsoft will continue to use OpenAI models through its partnership.

Microsoft is also developing its own model family from pretraining onward, Nadella says. He wants models specialized for areas such as security, knowledge work and coding that are efficient both to train and to deploy.

In a workflow that sends high-level intentions to multiple subagents, he expects different model families to be selected for different tasks. Microsoft’s stated goal, he says, is to provide choice rather than require one family for everything.

For engineers trying to become effective in the new era, Nadella recommends learning what is actually changing rather than following headlines and hype. His advice applies to graduates, mid-career engineers and CEOs alike: do more, learn hands-on and spend time with teams doing the work.

Orosz closes by drawing out the shift toward planning and validation, the need for companies to control their AI infrastructure, and the emergence of roles such as forward-deployed engineer. He expects reorganization and new engineering challenges to extend beyond Microsoft.

Más de este canal

1:25:20EN
The Pragmatic Engineer· 10 nov 2026Pódcast · 2 vocesSean Allen: Make software correctness part of the designInvitado: Sean AllenThe Antithesis engineer argues for earlier guarantees, more thoughtful test workloads and planning for production failures.Flujos de desarrollo y pruebasSoftware engineeringSeguridad+1
1:19:29EN
The Pragmatic Engineer· 5 nov 2026Pódcast · 2 vocesScaling Slack and GitHub: Maude Lemaire on Load, Shards and Reliable PlatformsInvitado: Maude Lemaire · Presenta: Gergely OroszThe principal engineer explains a Slack data migration, realistic load testing, GitHub’s shard constraints and her approach to AI-assisted performance work.Infraestructura y despliegueSoftware engineeringObservabilidad y fiabilidad+1
1:41:17EN
The Pragmatic Engineer· 28 oct 2026Pódcast · 2 vocesTitus Winters on Software Engineering, Teams and CS EducationInvitado: Titus Winters · Presenta: Gergely OroszA conversation on Google-scale refactoring, testing, AI and how engineers learn.Company BuildingLeadership
1:56:28EN
The Pragmatic Engineer· 21 oct 2026Pódcast · 2 vocesWill Larson: The Revised Rules of Engineering ManagementInvitado: Will Larson · Presenta: Gergely OroszThe 2010s taught engineering managers never to make a technical decision again. Will Larson, CTO of Imprint and author of Staff Engineer, argues that era is over: managers who cannot get back into the work are in a career trap.ManagementLeadershipCareer Growth+2
1:34:17EN
The Pragmatic Engineer· 14 oct 2026Pódcast · 2 vocesInside Uber’s Developer Infrastructure for AI AgentsInvitado: Uday Kiran Medisetty · Presenta: Gergely OroszUday Kiran Medisetty explains gateways, context, background coding and review.Flujos de desarrollo y pruebasIngeniería de agentesInfraestructura y despliegue
1:59:04EN
The Pragmatic Engineer· 7 oct 2026Pódcast · 2 vocesSam Newman on Resilient Distributed SystemsInvitado: Sam Newman · Presenta: Gergely OroszMicroservice tradeoffs, observability, retries and preserving understanding with AI.Software engineeringObservabilidad y fiabilidad