Microsoft’s AI Agents Now Decide What Code to Write

Microsoft is charting a bold new course for enterprise technology, shifting from AI assistants that follow commands to fully autonomous agents capable of deciding what code to write and assembling complete solutions. This evolution was a central theme at Microsoft Ignite 2025, where the company demonstrated how these agents are poised to fundamentally reshape software creation and operational management.

While the operational potential for self-monitoring and self-repairing platforms is significant, the advancements in autonomous software development are particularly striking. We are entering an era where AI won’t just assist with coding tasks but will take the initiative to determine what needs to be built and then construct those solutions from available components.

It’s important to approach this with a degree of caution. Current experiences with AI coding tools, while impressive in their capabilities, often involve a messy process. For every successful output, there can be numerous iterations where the AI misinterprets instructions, makes incorrect claims about its abilities, or produces unusable code. This highlights a crucial reality: even as these systems grow more powerful, they will require substantial human supervision for the foreseeable future. The notion of AI agents supervising other AI agents remains a distant prospect, underscoring the ongoing need for qualified human oversight.

A key announcement driving this vision is Microsoft Agent 365, a framework that grants a form of “personhood” to software agents. Drawing a loose analogy to corporate law, where companies are treated as legal persons, Microsoft is extending its user management infrastructure to govern these AI agents. They will be onboarded, managed, permissioned, and audited not as simple scripts, but as digital workers with unique identities. This shift enables agents to operate with intent, state, and context, moving beyond the predefined, deterministic tasks of traditional automated processes.

The practical mechanism enabling this autonomous assembly is the integration of the Model Context Protocol (MCP) into Microsoft Foundry. MCP acts as a universal translator, allowing different AI models and external services, like Slack, Google Drive, or Salesforce, to communicate through a standardized interface without requiring custom API integrations for every connection. This turns various software services into interoperable building blocks. Microsoft’s Foundry platform launches with a catalog of 1,400 MCP-compatible systems, empowering agents to dynamically snap together the tools they need to fulfill their objectives, effectively becoming mashup artists of enterprise software.

For agents to assemble solutions intelligently, they must move beyond simple data access to genuine understanding. This is the role of Microsoft’s new IQ services: Work IQ, Fabric IQ, and Foundry IQ. These services provide agents with the contextual awareness, semantic understanding, and long-term memory necessary for informed decision-making. Work IQ gives agents insight into how work flows through Microsoft 365. Fabric IQ provides business-level data meaning through semantic models. Foundry IQ offers unified knowledge access and persistent recall across diverse data sources. Together, they create a substrate for shared organizational understanding, allowing agents to comprehend not just what the data is, but what it means in a specific business context.

It’s clear that fully autonomous, self-building software environments are not yet a present-day reality. Microsoft’s announcements, however, outline the essential architectural components needed to make this future possible, with many features already available for preview. Progress will undoubtedly be incremental, often messy, and will continue to demand significant human guidance.

What are your thoughts on this direction? Do you believe agent-driven software assembly will redefine how organizations build and manage applications? Is this a practical near-term development or a longer-term ambition? How much human oversight do you envision these systems will require as they evolve?

(Source: ZDNET)