At Build 2026, Microsoft unveiled MAI, its family of LLMs, an agentic GitHub Copilot, and its Majorana 2 quantum chip

The Microsoft Build 2026 conference focused on developers designing agent-based systems, showcasing innovations in AI capabilities and agent execution on Windows 11. Major highlights include the introduction of a proprietary family of models called MAI, a reasoning model, the reimagined GitHub Copilot as an agent development environment, two open-source projects aimed at agent security, and the unveiling of a new quantum chip, Majorana 2.

Building an AI application has never been faster, shifting the challenge to ensuring the reliable operation of functioning systems rather than merely creating prototypes. As foundational inference tools become mainstream, the technology stack transitions away from conversational assistants toward agents. Microsoft has made this evolution the central theme of its annual developer conference.

The company aims to address the fragmentation in the AI landscape by providing technical teams with tools to design, execute, and deploy end-to-end agent systems while retaining control over their execution environment. Announcements from Build 2026 span the entire spectrum, from developer workstations to agent governance mechanisms, venturing into unexpected territory with a significant leap into quantum computing.

Redefining the Developer’s Environment for Local Execution

The first aspect targets Windows 11, which Microsoft aspires to position as the execution environment for AI and agents. The company is expanding its Windows AI APIs, integrating agent execution environments like OpenClaw, and introducing specialized hardware such as the Surface RTX Spark Dev Box. The stated goal is to enable developers to run demanding AI and agent workloads locally and at scale.

This emphasis on local execution is significant for technical leads. Running agents on-site rather than in the cloud addresses concerns around latency, inference costs, and data privacy. This shift distributes some of the workload to the developer’s hardware, which requires adequate equipment but allows for testing and operating agent chains without exposing data to remote services. For organizations keen on controlling their information flow, this local execution capability is a compelling advantage.

A Proprietary Family of Models

Microsoft has confirmed its commitment to developing its own models with the MAI family. At the forefront, MAI-Thinking-1 is a reasoning model comprising 35 billion active parameters designed for complex tasks and code generation, currently available in private preview on the Microsoft Foundry platform. It is accompanied by a broader multimodal family that covers imagery, voice, transcription, and coding, available immediately.

This expansion of in-house models highlights the company’s trajectory. Once reliant on OpenAI’s models, Microsoft is developing its own capabilities, granting it greater flexibility regarding costs, specialization, and control over its stack. For teams building on Foundry, the introduction of a proprietary reasoning model broadens their options while still raising questions about dependency on the ecosystem. A proprietary model, even when offered by its developer, remains a component whose availability, pricing conditions, and evolution are outside the client’s control.

GitHub Copilot Transitions to Agent Functionality

The most significant announcement regarding development practices centers on GitHub Copilot, now envisioned as an application designed within an agent development environment. It’s rooted in GitHub’s collaborative graph, alongside models and workflows that developers already utilize. This terminological shift is noteworthy. An assistant offering suggestions under continuous supervision can be termed a co-pilot, while a system executing sequences of actions without human validation qualifies as an autonomous agent, with substantial implications for responsibility and traceability.

This evolution is not merely a productivity gain for technical and legal teams. When an agent interacts with a code repository, opens tickets, modifies files, or executes operations without intermediate human checkpoints, responsibility for those actions shifts from the user to the agent operator. Promised time savings, often touted by tool providers, must be assessed in light of this transition and the safeguards it necessitates in continuous integration and delivery pipelines.

Two Open-Source Foundations for Agent Governance

Microsoft is backing this movement with two open-source projects focused on agent security. The first, named ASSERT, is designed to evaluate an agent’s security compliance with defined governance policies through a rating and regression testing mechanism. The second, Agent Control Specification, aims to standardize application points and control mechanisms within agent execution loops.

This focus on control addresses growing concerns among security leaders. An autonomous agent expands an organization’s exposure surface, interacting with tools, processing data, and triggering actions. Providing frameworks to evaluate agent behavior against explicit rules and to enforce control points in execution loops offers the necessary elements to support agent governance. The choice of open-source enables third-party examination of these mechanisms, fostering transparency for components aimed at regulating autonomously acting systems.

A Claimed Quantum Milestone

The final component extends beyond software development. Microsoft unveiled Majorana 2, its latest quantum chip, claiming a reliability improvement of a thousandfold over its predecessor. The company asserts that this advancement paves the way for a scalable quantum computer by 2029, halving its original timeline. Such claims are yet to be confirmed through independent publications and validations, positioning the quantum race within a landscape that several competitors are vying to dominate.

The company links this announcement to the general availability of Microsoft Discovery, an AI agent solution aimed at accelerating scientific research by executing iterative workflows alongside human expertise, now accessible to customers in local preview. The combination of a quantum milestone and an agent platform for research delineates Microsoft’s ambitious vision for scientific computing. However, for decision-makers, the 2029 target and the announced reliability factor remain aspirational milestones, to be monitored closely given the history of shifting timelines in the quantum computing arena.

Scroll to Top