Microsoft makes its Copilot harness available through Foundry

Microsoft says Foundry provides a fully managed way for developers and enterprise customers to build agents that reason, use tools and handle complex, long-running tasks.[1] The newly available Copilot harness is the orchestration layer that gives an agent its instructions, context, execution environment and controls.[1] Microsoft says the harness already powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI and VS Code.[1] The company says its platforms will remain open to other harnesses and frameworks.[1]

Microsoft says the harness supports more than 20 frontier models, bring-your-own-key use, and routing work to models suited to a task.[1] It says customers can choose a model without switching the harness or the rest of their technology stack.[1] Microsoft operates the agent loop and sandbox, with governance and optimization built in; customers provide the instructions, skills, tools and model choice.[1] The harness supports open standards such as MCP and agent skills, as well as agent plugins.[1] Microsoft says model choice matters because a frontier model today might fall behind tomorrow.[1]

Unite.AI reports that preview agents can work with files and code, use connected Foundry and MCP tools with configured authentication and approvals, and delegate bounded tasks to built-in subagents.[2] The report says developers can author agents through the Azure Developer CLI preview surface, Foundry portal, Foundry Toolkit for VS Code, or a GitHub Copilot app plugin and skill.[2] Agent definitions can be kept alongside application code under source control, according to the report.[2] One preview limitation is that customers select the Copilot harness when creating an agent and cannot yet switch harnesses afterward.[2]

Unite.AI reports that long-running resilience for hosted agents is generally available, while the Copilot harness itself is in public preview.[2] Foundry detects interruptions and restarts hosted-agent work with its original input; customer code must save progress and intermediate results for resumption.[2] Workflows that update another system also need to prevent duplicate actions when work restarts.[2] The report says harness runtime usage is free before a specified date, but model inference and separately billed tools and connected services remain chargeable.[2] Starting December 1, sandbox compute is reported at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, with prices varying by region and agreement.[2]