Picking an AI model can feel a bit like buying a smartphone just before the next one gets announced. You choose something that looks capable, only to see another company arrive with fresh claims about better reasoning, faster responses, or lower costs. For businesses building AI agents, that creates a practical question: how much work will it take to change models later?
Microsoft is pitching its new Copilot harness as an answer. This technology gives developers a shared foundation for running agents across different models, with Microsoft handling much of the software that turns an AI response into actions. The company is also making it available through Microsoft Foundry.
“Harness” is another term the AI industry now expects everyone to understand, but the basic idea is straightforward. A model supplies the intelligence, while the surrounding software manages instructions, context, tools, and the environment in which an agent works. Having a capable model is useful, but getting it through a task involving multiple steps requires more than a chat box.
The technology originated with GitHub’s engineering team, and Microsoft says it now powers experiences including Copilot Studio, the GitHub Copilot app, the CLI, and VS Code. The company describes support for more than 20 frontier models, along with bring-your-own-key access. It also says it recently moved the harness runtime to Rust, claiming reductions in session overhead and time to first token.
For developers, the concrete announcement is a public preview of the Microsoft Copilot harness in Foundry Agent Service. It combines reusable skills, connected tools, file and code execution, and the ability to delegate bounded tasks to subagents. Developers supply the instructions and capabilities, while Microsoft manages the agent loop and sandbox.
Think about an agent preparing a research brief. It might need to gather information, analyze a spreadsheet, check conflicting claims, and produce a document with citations. The appeal of a managed system is that developers can spend more time defining what good work looks like and less time assembling the machinery that moves between those steps.
That does not remove the need to check the finished work. An agent can successfully operate its tools and still deliver a report with a bad conclusion, which is an important distinction when companies start presenting task execution as evidence of intelligence.
Microsoft’s own figures offer an interesting angle here. In internal Copilot Studio evaluations, changing the harness improved pass rates by 10.8 percentage points with an OpenAI model and 22.1 percentage points with an Anthropic model. The company says changing models had comparatively little effect in those tests.
Those results deserve attention, but they are Microsoft’s internal findings. They do not establish that its approach will deliver the same improvement for every business, and they certainly do not settle which model is best. What they suggest is that the software around a model deserves scrutiny alongside the model itself.
There are practical limitations to the Foundry preview, too. Switching the harness after creating an agent is not currently supported, and Microsoft manages the execution environment’s CPU, memory, and idle settings. Developers get convenience, but they should understand the boundaries before building around it.
Then there is the bill. Harness runtime usage is free before December 1, 2026, although model inference, separately billed tools, and connected services remain subject to their charges. Consumption-based billing for the harness begins on that date.
To me, the bigger story is Microsoft’s position in the AI business. If companies keep switching between competing models, Microsoft still has an opportunity to sell the infrastructure that runs their agents. Model choice can be good for customers while also being good for the company collecting the infrastructure revenue.
That makes this more interesting than another Copilot branding announcement. Microsoft is making a case that the system surrounding AI matters as much as the model powering it, and the useful test will be whether customers can complete dependable work at a cost they can justify.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz
