Microsoft Opens Copilot Agent Harness to Builders via Foundry

Microsoft has opened the orchestration engine behind its Copilot products to outside developers. According to reporting by The New Stack, the GitHub Copilot agent harness is now available in public preview through Foundry Agent Service, giving builders a managed way to run autonomous agents without assembling the plumbing themselves.

The move brings the same Copilot agent harness infrastructure that powers Copilot Code, Copilot Studio, GitHub Copilot, the command-line interface, and Visual Studio Code into a service any developer can use. Microsoft announced the release on October 28, and The New Stack published its account of the announcement on October 29. The launch continues the Foundry strategy Microsoft laid out at its Build 2026 conference, positioning the cloud platform as a home for agent development as well as model hosting.

What the Copilot agent harness actually does

The Copilot agent harness handles the work that sits between a model and a finished task. It manages model calls and tool calls, tracks conversation context, and keeps an agent operating within the policies its developer configured. Built-in sandboxing isolates the agent's execution environment, while reusable skills and integrations with the Model Context Protocol let agents connect to outside tools and data sources. Support for subagents means a main agent can delegate parts of a job to smaller, specialized agents working under its direction.

Microsoft runs the agent and manages its execution environment, so developers focus on choosing a supported model and configuring instructions and tools. The New Stack's walkthrough of Microsoft's own technical example shows how that looks in practice: a research agent defined by a Markdown-based skill that sets out steps for collecting evidence, analyzing data, and writing a report, with an azure.yaml file defining the model deployment, tools, instructions, and harness type. Developers select microsoft_copilot_preview as the harness type, deploy with the Azure Developer CLI, and keep their configurations in source control.

That managed approach is the pitch of the Copilot agent harness. Building a reliable agent usually means solving the same problems every team solves separately: retrying failed tool calls, trimming context before it overflows, containing an agent that goes off track, and auditing what it did. The Copilot agent harness folds those concerns into the platform, which could shorten the path from a promising prototype to a production deployment.

The trade-offs developers should weigh

The constraints are where The New Stack centers its skepticism, and they deserve attention from any team considering the service. The harness choice must be made when a prompt agent is created, and it cannot be changed later, even in a new version of the agent. That lock-in means an early architecture decision carries long-term weight, and teams should prototype carefully before committing a production agent to the Copilot harness.

Microsoft also controls the compute envelope: it sets how much CPU and memory an agent receives and how long it can sit idle. Developers pick from compatible models deployed through Foundry, but bringing their own models is not supported with the managed harness. The New Stack reads that limitation as a tension with Microsoft's multi-model positioning, since teams that have standardized on a model outside the supported set cannot simply plug it in.

Teams that want more control have another route. Foundry's hosted agents let developers run their own frameworks while Microsoft handles hosting, and long-running resilience for those hosted agents reached general availability on the same day as the harness preview. The choice, in effect, is between convenience and control: the Copilot agent harness for teams that want Microsoft to run the loop, or hosted agents for teams that want to own it.

Why this matters for the agent ecosystem

The launch lands in a week crowded with agent infrastructure news. Enterprises are racing to give agents governed access to their systems, as shown by recent coverage of agent platforms for financial firms and by new reporting on enterprise AI spending. Identity, observability, and security layers for agents are all attracting investment, and the harness is the execution layer those tools will increasingly wrap around.

For Microsoft, the strategic logic is clear. Copilot products already reach millions of users, and the Copilot agent harness turns that internal capability into a platform play. Every developer who builds on Foundry Agent Service deepens the company's position in the agent stack, from model hosting through orchestration to the end-user products where agents do their work. If the preview proves reliable, it could become a default choice for teams already inside the Microsoft ecosystem.

For the broader field, the release raises the bar on what a managed agent service should include. Sandboxing, skills, subagents, and protocol integrations are becoming table stakes rather than differentiators, and the competition now shifts to flexibility: which models a service supports, how much control it hands back to developers, and what it charges once the free period ends.

Pricing and what comes next

The Copilot agent harness in Foundry Agent Service is free until December 1, 2026, though model inference and some tools and connected services carry separate charges. After that date, Microsoft will charge $0.0994 per vCPU-hour for compute and $0.0118 per GiB-hour for memory, with rates varying by region and customer agreement, according to The New Stack's account of the announcement.

The preview period gives developers roughly a month to test the harness without paying for its compute. The real test will be what teams build in that window, and whether the convenience of the Copilot agent harness outweighs the lock-in of an unchangeable harness choice. Watch for early production stories, independent benchmarks on reliability, and any expansion of the supported model list, which would soften the most notable limitation of the launch.

Sources: reporting by The New Stack on Microsoft's Foundry Agent Service announcement; Microsoft's Build 2026 Foundry strategy coverage.