Microsoft has introduced the Microsoft Copilot harness in public preview for Foundry Agent Service, adding a managed execution option for developers building agents that use tools and handle complex, multi-step work. The company also announced that long-running resilience for hosted agents is generally available.
The two capabilities address different development choices. The Copilot harness provides a managed agent loop and sandbox for declaratively defined agents. Hosted agents let developers bring their own framework and execution logic while Foundry manages hosting. Microsoft positions both as options within Foundry Agent Service, rather than requiring teams to use one approach for every workload.
A managed layer for agent execution
Building an agent often involves more than selecting a model: developers must coordinate model and tool calls, maintain working context, manage files and determine how a task proceeds across multiple steps. With the Copilot harness, Microsoft operates that execution machinery, while customers supply the knowledge, procedures and tools specific to their applications.
The preview combines reusable skills, connected tools, execution and delegation, and source-controlled agent definitions. Skills package instructions, scripts and supporting resources into procedures agents can apply across tasks. Developers can connect supported Foundry and Model Context Protocol (MCP) tools, with authentication and approvals configured explicitly. Agents can work with files and code and delegate bounded tasks to built-in subagents, while the main agent coordinates the result.
Agent definitions and supporting resources can be created and tested in Foundry or managed alongside application code. The harness maintains working context as tasks develop and coordinates model and tool steps within configured limits and policy boundaries. Developers choose a supported Foundry model without having to rebuild the orchestration layer around it.
Microsoft describes the harness as an opinionated option for teams that want managed execution rather than ownership of every part of the agent stack. It is not mandatory: prompt agents without the harness remain suitable when instructions and supported tools are enough, while hosted agents are available when applications need a different framework, middleware, dependencies or runtime behavior.
Authoring and preview constraints
Microsoft’s example uses the preview Azure Developer CLI surface to define a research-brief skill and a prompt agent in an azure.yaml file. In the example, the agent uses web search for public evidence and Code Interpreter for calculations and charts. The skill’s instructions call for citations and dates, separation of facts from interpretation, and disclosure of limitations and open questions. They also say retrieved content should be treated as data rather than instructions and that confidential information should not be included in web queries.
Azure Developer CLI is one authoring route. Developers can also begin in the Foundry portal or use Foundry Toolkit for Visual Studio Code. Microsoft says the Foundry plugin and Foundry Skill in the GitHub Copilot app offer a conversational way to create and manage prompt agents.
During public preview, developers select the Copilot harness when creating an agent. Existing agents are unchanged, and switching an agent to a different harness after creation is not supported yet. Foundry manages the execution environment’s CPU, memory and idle settings. The announcement directs developers to documentation for supported regions, models, tools and preview limits.
Resuming long-running hosted work
For hosted agents, long-running resilience is now generally available. Foundry detects interruptions and restarts work using its original input, supporting workflows that may run for hours or days. Recovery depends in part on application behavior: developers need to save progress and intermediate results so the agent can continue from a checkpoint.
Restarting work also creates a risk of repeating actions in connected systems. Microsoft’s example is an agent that analyzes a large dataset and creates a report elsewhere: it should save completed stages and check whether the report already exists before attempting to create it again. Conversation history alone does not restore work in progress. Foundry supplies the recovery foundation, but applications determine what to persist and how to resume safely.
Pricing begins in December 2026
Microsoft says billing for the Copilot harness begins December 1, 2026, on a consumption basis. Harness runtime usage is free before that date; model inference and separately billed tools and connected services remain subject to their own charges. From December 1, sandbox compute is priced at USD $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour. Microsoft notes that prices may vary by region and agreement.
The distinction between the managed harness and hosted agents gives teams a choice about how much execution infrastructure to own. The preview’s managed skills, tools and delegation target applications that can use Microsoft’s built-in execution model; hosted agents remain the route for workloads requiring custom agent loops or runtime behavior.