Microsoft launches Copilot harness for Foundry Agent Service
Microsoft is offering its internal Copilot runtime as an optional tool for developers to build long-running agents with built-in skills.
The short version
- The harness powers existing products like GitHub Copilot, VS Code, and the Copilot CLI.
- It supports over 20 frontier models and includes reusable skills and file/code execution.
- Microsoft manages the approval policy for built-in tools, with shell access being on by default.
- The runtime is free until December 1, 2026, after which sandbox compute and memory will be billed.
- Unlike AWS, Microsoft has not yet published a way to export harness agents into framework code.
Why it matters
This move makes it easier for companies to deploy complex AI agents without building the underlying 'loop' from scratch. However, because Microsoft controls the tool approvals, IT leaders must balance ease of use against security oversight.
A shared runtime for agents
Microsoft recently introduced the Microsoft Copilot harness as an optional runtime within the Foundry Agent Service. This harness is the same underlying technology that powers several of Microsoft's major AI products, including GitHub Copilot, Copilot Studio, and the Copilot command-line interface. By making it available in Foundry, Microsoft is allowing customers to leverage a production-ready framework for building prompt agents.
The harness is designed to handle the complex logic required for agents to reason, use tools, and complete long-running tasks. While the customer provides the specific instructions, skills, and models, Microsoft manages the agent loop and the sandbox environment. This means developers do not have to build the foundational infrastructure for how an agent interacts with its environment or handles multi-step processes.
As reported by The Futurum Group, the harness supports more than 20 frontier models and allows for bring-your-own keys. It bundles several components together, such as reusable skills, connected tools with configured authentication, and file or code execution capabilities. These features are intended to help teams build sophisticated agents that can perform actions rather than just generating text.
Security and approval limits
One of the most significant characteristics of the Microsoft Copilot harness preview is how it handles tool approvals. Microsoft currently manages the approval policy for the harness's built-in tools, which means the platform sets the permissions for these capabilities. New agents are granted access to all five built-in tools by default unless a team specifically chooses to turn them off.
The documentation highlights that the shell is considered the broadest harness tool available to users. Because of this, Microsoft warns that disabling certain features like file or web access while leaving the shell active does not create a strict security boundary. Teams can use network egress controls to limit outbound requests, but the underlying approval model remains largely under Microsoft's control.
This differs from some other offerings where users might have more granular control over the core framework. For example, the service manages the approval settings for Model Context Protocol (MCP) tools that a team configures. This creates a specific trade-off where customers gain a faster path to deployment in exchange for less direct authority over the harness's native tool permissions.
Comparison with competitors
Microsoft is entering a space that already has established players, most notably Amazon Web Services (AWS). AWS released its AgentCore harness in mid-2026, which allows customers to switch model providers mid-session and export agents to code with a single command. In contrast, Microsoft has not yet published a similar export path from the Copilot harness to framework code.
Anthropic also offers Claude Managed Agents, but those are currently limited to supporting only Claude models. Microsoft’s offering is broader in terms of model variety, supporting over 20 different frontier models. However, the lack of an export tool means that teams who choose the Copilot harness are making a more permanent commitment to the Microsoft runtime than they might with AWS.
For organizations that need to own the entire agent loop, Microsoft provides hosted agents as an alternative. These hosted agents run on the GitHub Copilot software development kit (SDK) without requiring rewrites. However, moving away from the harness means the customer becomes responsible for preserving progress and preventing duplicate side effects during interrupted tasks.
Costs and production readiness
The Microsoft Copilot harness is currently in public preview, which means it does not come with a service-level agreement. Microsoft advises that it is not recommended for production workloads at this stage. The runtime itself is free until December 1, 2026, but users will still be responsible for model inference and other separately billed services.
Starting in December, Microsoft will begin billing for sandbox compute and memory. The rates are set at $0.0994 per vCPU-hour and $0.0118 per GiB-hour, though these prices can vary depending on the region or specific agreement. These costs are added on top of the existing charges for the models and tools the agents use to perform their work.
Because the choice of the Copilot harness is fixed once an agent is created, platform leaders need to consider the long-term implications of this choice. While the harness simplifies the initial build, the lack of an export path means that moving an agent to a different framework later would likely require a significant rewrite. This makes the initial decision a long-term architectural commitment for the organization.
Strategic implications for IT
For Chief Information Officers and platform engineering leaders, the Copilot harness represents a shift in how AI agents are governed. Since the choice of runtime is made at the moment of creation, the decision often moves from formal architecture reviews to the individual developers building the agents. This can lead to a situation where agents are deployed using a runtime that may not perfectly align with corporate security policies.
The approval split is a key area of concern for auditors and regulators who need to know exactly what an agent can reach. As agents move from suggesting content to executing actions, the ability to control those actions becomes a bottleneck for adoption. If a team cannot set specific approval controls for built-in tools, they may face challenges during security audits or incident reviews.
To manage this, platform teams should encourage developers to keep the layers above the harness portable. By using open standards like the Agent Skills format and Model Context Protocol (MCP), teams can ensure that their skills and servers remain reusable. This approach helps mitigate the risk of vendor lock-in while still taking advantage of the high-level capabilities provided by the Microsoft Copilot harness.
What to watch
- Whether Microsoft adds a code export path or customer-controlled tool approvals before billing starts.
- If Anthropic opens Claude Managed Agents to other model providers to compete with multi-model harnesses.
- New task-completion results from Foundry customers once the harness moves out of preview.
Try it
Check if your team's current AI agents use a fixed runtime or if they can be exported to standard code frameworks.
Sources and links
- OfficialAnthropic’s Claude Managed Agents platform.claude.com
- OfficialTina Schuchman azure.microsoft.com
- OfficialMicrosoft’s Foundry announcement of the Copilot harness devblogs.microsoft.com
- OfficialJune 2026 Build recap devblogs.microsoft.com
- Officialimplementation guide learn.microsoft.com
- OfficialA companion post blogs.microsoft.com
- ReportingThe Futurum Group: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by The Futurum Group · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
