Zuckerberg and Nadella on AI Coding, Agents and the Next Software Stack

Matthew Berman reviews excerpts from a LlamaCon 2025 conversation between Meta CEO Mark Zuckerberg and Microsoft CEO Satya Nadella. He begins with their discussion of AI coding and then moves through research, agents, productivity and open models. The order is selected by Berman rather than a complete replay of the event.

Zuckerberg asks how much code at Microsoft is written by AI. Nadella says the company tracks acceptance rates and notes that results vary by language: in his account, Python support has been stronger than C++ support, while tools have also become useful for code review.

Berman suggests training data and tooling may contribute to those differences, though the video does not establish a cause. He argues that large organizations face dependencies and legacy systems that can make autonomous changes harder than work in a new, smaller project.

For code review, an assistant can focus on a change and its surrounding context. Berman sees promise in that narrower task. He also cautions that a metric for AI-written code may combine accepted autocomplete with code written through an agent, even though those represent different degrees of automation.

Nadella estimates that software wrote around 20–30% of code in some Microsoft projects. Zuckerberg says he does not have Meta's corresponding figure at hand and observes that many commonly reported statistics still concern autocomplete. The estimate is limited to the projects described, not all Microsoft code.

Zuckerberg mentions contained uses such as feed and advertising ranking experiments, then identifies a larger goal: building an AI machine-learning engineer to advance Llama itself. Berman calls this an especially consequential ambition and imagines a possible feedback loop in which AI helps improve future AI.

Zuckerberg speculates that perhaps half of development could be done by AI within the following year, with the share increasing afterward. Nadella separates tasks such as kernel optimization and security improvements from inventing a new model architecture, which he regards as less likely in that time frame. These are forecasts, not verified achievements.

Berman describes his own hypothesis of an intelligence explosion, but faster recursive improvement does not automatically follow from using AI in research. Zuckerberg distinguishes Meta's internal model-development goal from Microsoft's broad developer-tool business, while Berman speculates that the companies' roles may overlap more over time.

Nadella suggests rethinking infrastructure and tools for agents as users: what should their sandboxes, repositories and interfaces look like? Berman extends the question to database access, browsing, communication and standardization. He sees a software stack designed around agent use as a major open design problem.

Zuckerberg imagines engineers increasingly acting like technical leads working with groups of agents. Berman calls this orchestration and wonders how multiple agents might coordinate changes to a repository. He emphasizes uncertainty about the eventual arrangement rather than presenting a settled design.

For people starting in development, Berman still sees value in learning to code, especially because it teaches systems thinking. He predicts that interacting with AI agents may become more important than fluency in any one programming language. That is his personal view of a changing profession.

Nadella considers the boundary between a chat, a document and an application. An interaction can collect information into a persistent artifact and then become something more like an app. He suggests that some familiar software categories reflect old technical constraints rather than permanent distinctions.

Berman imagines a future in which data, agents and human oversight work together in a different interface. He is unsure whether today's separate applications would retain the same shape, and treats this as an open question rather than a product already delivered.

Asked how to judge AI's progress over several years, Nadella says investment must produce real productivity gains across functions and industries. He argues that better models alone are insufficient: software, management and the way people work also need to change.

Nadella compares adoption to the long process by which factories reorganized around electricity. His point is that simply inserting a new technology into an old workflow may miss its potential. Berman agrees that AI could help address difficult problems, while its actual economic impact remains to be shown.

Nadella describes each platform transition as a chance to reconsider the technology stack, including storage and infrastructure for training workloads. Berman compares the early web, when publishers initially copied print formats online, with today's tendency to add AI to familiar development tools.

Berman points to newer coding interfaces as early attempts to design around AI rather than just add it to an existing editor. He imagines humans coordinating agents and argues that the eventual interaction model is still being invented.

The conversation turns to efficiency and demand. Berman invokes Jevons paradox: lower unit costs can lead to greater total use, though whether use rises enough to offset efficiency gains depends on the circumstances. Nadella points to improvements across chips, systems, model architecture and inference as drivers of lower costs.

Berman sees opportunities both in infrastructure—memory, agent frameworks, tool use and prompt management—and in applications that deliver model capabilities to users. Nadella discusses workflows combining agents built on different models and the value of protocols and open models for interoperability.

Berman imagines routing individual tasks to models from different providers based on cost, efficiency and quality. He treats this as an emerging application pattern, not a guarantee that a particular routing strategy will work best in every case.

Zuckerberg asks about Microsoft's relationship with open source. Nadella says interoperability matters to customers and that both open and closed approaches have roles. He also describes enterprise interest in owning or distilling models; Berman responds that, in his experience, retrieval and context often meet business needs without fine-tuning.

On agents inside Microsoft, Nadella describes an evolution from code completion to chat and assigned tasks. He says these modes can coexist and must integrate with existing repositories and developer workflows, since established teams rarely work exclusively on new projects.

Berman is more excited by the possibility of designing a coding environment around agents from the start. He asks what an orchestration interface, or even code itself, might look like if agents write and read most of it. He acknowledges these are questions about a possible future.

Berman ends by citing a claim from Cursor's Aman Sanger about the daily volume of accepted AI-generated code and comparing it with an estimate of worldwide code production. These figures are claims in the video, not independently verified here, and accepted lines do not measure software quality or productivity. Berman expects natural-language tools to broaden who can build software.