Microsoft Build 2025: Satya Nadella goes all-in on the agentic web with MCP, NL Web, and xAI partnership

Summary

Satya Nadella positioned Microsoft as a model-agnostic cloud platform at Build 2025, culminating in a public partnership with Elon Musk's xAI. The move signals a notable shift between the two executives, who faced legal friction as recently as 2023 when Musk threatened to sue Microsoft over Twitter data used for model training and was also named as a defendant in Musk's OpenAI lawsuit.

The partnership integrates xAI's Grok 3 model into Azure AI Foundry, letting enterprise customers mix OpenAI, Meta's Llama, and xAI models without vendor lock-in. Microsoft's bet is that businesses will choose its platform precisely because they can avoid being locked into a single vendor. This positioning directly targets Google's strategy of pushing Gemini on cloud customers.

The agentic web

Microsoft's core announcement rests on a shift where agents rather than humans become the primary consumers of web services. The company introduced NL Web, a markup protocol that lets any website expose a conversational endpoint discoverable by AI agents. Copilot Studio now supports multi-agent orchestration, allowing agents to delegate tasks to one another.

This change breaks existing business models. If agents replace human traffic, ad-supported revenue collapses. Derek Thompson, citing Ben Thompson's analysis, describes a scenario where an agent books a child's birthday party, purchases flights, and edits work memos while the user sleeps. Ad revenue becomes meaningless when the reader is a robot. Thompson proposes micropayments or blockchain-based stablecoins as replacement revenue, though conventional dollars could work too.

Consumer tech companies like Instacart, Uber, and Apple eventually become advertising businesses because ads carry high margins. If agentic software strips out that ad layer, it removes earnings. The challenge is structural resistance from platforms with strong ad businesses that won't cede control to agents bypassing display revenue.

One counterargument exists: if LLMs train on advertising content embedded in podcasts, articles, and videos, future models will encode brand preferences from their training data. An agent simulating human behavior would then simulate susceptibility to ads.

Coding agents and market fragmentation

GitHub Copilot graduated to a full autonomous coding agent, Project Padawan, available for Copilot Enterprise and Pro Plus users. It handles autonomous feature work, bug fixes, and refactors. The coding agent market now includes three OpenAI products (o3, o4, and CodeGPT), Cursor's Windsurf, and Cognition's Devon, each targeting different user profiles.

Four distinct markets exist: consumers who don't realize they're writing code (code generated transparently), "proumers" using CodeGPT to modify GitHub repos without opening an IDE, professional programmers using Windsurf as an IDE copilot, and top-down enterprise deployment like Devon. Caleb Harris posted an example of the proumer angle, dropping 12 backlog bugs into CodeGPT, having Graptile catch nits, and running tests through Spur QA with almost entirely human out of the loop.

The market opportunity is massive given the scale of software engineering spend, but nascent. Unlike coding IDEs or CI/CD tools, there's no established spend to displace. The revenue is purely additive, making long-term competitive positioning difficult to predict.

Other announcements

Copilot Tuning lets companies fine-tune models with proprietary data to build domain-specific agents. Windows AI Foundry launches as a local platform on Windows and Mac for training and fine-tuning LLMs. Model Context Protocol (MCP) receives first-party support across Windows and Azure. SQL Server 2025 entered public preview.