Microsoft Corp. today announced it’s bringing agentic development capabilities into the hands of developers with a new desktop form factor supercomputer called the Surface RTX Spark Dev Box.

Developed with Nvidia Corp.’s newly announced RTX Spark processor under the hood, it’s capable of delivering up to 1 petaflop of AI compute with 128 gigabytes of unified memory and run up to 120 billion-parameter models. — as long as they’re crunched down to 4-bit quantization, a type of compression that allows larger models to run on tighter hardware.

As a Windows box, it’s capable of running them on Microsoft’s Linux virtualization WSL 2 with native graphics processing unit passthrough with full CUDA support preconfigured for developers with Visual Studio Code, GitHub Copilot and other tools preinstalled.

The company said it would become available later this year in the United States.

Adding isolation and control to agents

Although Microsoft began with announcing hardware, the company continued to innovate around software, especially on the ground level: the operating system. The company said it wants to build Windows as an “agent-native runtime.” To do this, the OS now has Microsoft Execution Containers, or MXC, now in preview, which give developers and information technology administrators the ability to create enterprise-grade sandboxed environments to isolate agents.

This technology is now being used by OpenClaw on Windows, a port of the now ultra-popular personal open-source agent, which allows multistep autonomous agentic actions. Using MXC, the OS itself creates sandboxed boundaries around the agent that manages inference, executes policies, obfuscates personally identifying information and limits the agent’s visibility of the system and data.

The same model is available for agents in the cloud with hosted agents in Foundry Agent Service, in preview. It’s an instant-on system for sandboxing agents at scale in the cloud, per session, isolated execution, persistent memory and elastic scale.

Empowering developers with new tools

Microsoft is bringing the GitHub Copilot agentic development experience directly to the desktop with a native application, the GitHub Copilot app, now in preview.

Developers can start with an idea, written in plain language, an existing issue or a proposal to merge code. They can then orchestrate multiple agents in parallel and keep changes moving in review, integration and merge the final output.

Each session uses git worktrees, a native git feature that allows developers to check out and work on multiple branches simultaneously in separate directories, all linked to a single history without colliding. Normally, a git repository only allows users to have one branch at a time; users who want to switch must either commit unfinished code or use “git stash.” Copilot handles this execution, while developers maintain control.

For scientists, Microsoft announced that Discovery is generally available today. This platform provides researchers with an enterprise-grade agentic AI platform for science workflows built on Azure. It helps automate research and development using specialized agents that can reason over public-domain knowledge, create hypotheses on vast search spaces, test and validate hypotheses at scale, analyze results and return conclusions through iterative loops.

BHP Group, Syensqo SA/NV and GlaxoSmithKline plc have been using the system to accelerate research and development across the scientific community.

A free Discovery local app was announced for the broader scientific community as well. That will be available this summer in preview and only requires a GitHub Copilot account.

Image: SiliconANGLE/Microsoft Designer