Particle.news

Microsoft Pushes Agent-First Stack With Solara, Scout and MAI Models

The company says the package will let enterprises run persistent AI agents on devices under OS-enforced sandboxes to cut reliance on outside model providers.

Overview

  • Microsoft unveiled a coordinated agent-native roadmap at Build 2026 on Tuesday that includes Project Solara (an Android-based agent platform), Microsoft Scout (an always-on Autopilot agent), a new MAI model family, developer hardware and OS-level execution containers.
  • Microsoft Scout is built on the OpenClaw framework and WorkIQ context engine and is available now to Frontier customers as a limited preview that will be rolled out more broadly later.
  • MAI-Thinking-1 is presented as Microsoft’s first reasoning model with about 35 billion active parameters and a 128K context window, and the MAI family is being delivered through private and developer previews rather than wide commercial release.
  • For on-device work, Microsoft showed the Surface RTX Spark Dev Box — a developer desktop using Nvidia’s Spark chip with up to 128GB unified memory — and demonstrated running very large models locally, while Project Solara was shown as prototype reference devices for partner pilots, not consumer products.
  • Microsoft introduced Microsoft Execution Containers and identity controls to enforce sandboxing of agents on Windows, framing the announcements as a push to own more of the AI stack and give enterprises control, though most components remain in previews and broad adoption is still unproven.