Microsoft announced the Microsoft Copilot harness in Foundry Agent Service on 28 October 2026, making the managed agent execution layer available in public preview. The release is aimed at developers building agents that need to coordinate model calls, tools, files and multi-step work without owning the full orchestration stack.
Microsoft also announced general availability of long-running resilience for hosted agents. That is a separate capability for teams that bring their own framework, agent loop, middleware, dependencies and runtime behaviour, while using Foundry to host the agent.
What the harness provides
The Copilot harness runs the agent loop and sandbox in Foundry Agent Service. Developers define the agent declaratively, choose a supported Foundry model, and provide the knowledge, procedures and tools that the agent can use.
The preview has four main building blocks:
- Reusable skills, which package instructions, scripts and supporting resources into procedures an agent can apply across tasks.
- Connected tools, including supported Foundry and Model Context Protocol tools, with configured authentication and approvals.
- Execution and delegation, including work with files and code, and bounded tasks delegated to built-in subagents.
- Source-controlled agent definitions, so teams can create and test in Foundry or manage definitions with application code.
The harness manages working context as a task develops and coordinates model and tool steps within configured limits and policy boundaries. Microsoft positions this as the managed option between simple prompt agents and hosted agents where the application team owns more of the runtime behaviour.
Authoring options
Developers can start in the Microsoft Foundry portal or use the Azure Developer CLI. The example configuration in Microsoft’s post uses the preview harness type microsoft_copilot_preview, with skills, instructions, tools and container settings described in azure.yaml.
Foundry Toolkit for Visual Studio Code is another authoring route. Microsoft also lists the Microsoft Foundry plugin and Microsoft Foundry Skill in the GitHub Copilot app as a conversational way to create and manage prompt agents.
During the preview, developers select the Microsoft Copilot harness when creating the agent. Existing agents are unchanged, switching the harness after creation is not yet supported, and Foundry manages the execution environment’s CPU, memory and idle settings.
Hosted agents and recovery
Hosted agents remain the option for teams that need to bring their own framework or agent loop. With long-running resilience, Foundry can detect an interruption and restart the work with the original input.
That recovery support does not remove application responsibility for state. Microsoft says the agent’s code must save progress and intermediate results so it can resume from a checkpoint. If the workflow writes to another system, the application must also prevent duplicate actions when work restarts.
The distinction matters for workflows that run for hours or days. Conversation history alone is not enough to restore work in progress, and a restarted agent may otherwise repeat a ticket creation, report generation or similar external action.
Pricing and preview limits
Billing for the Copilot harness in Foundry Agent Service starts on 1 December 2026. Harness runtime usage is free before that date, but model inference, separately billed tools and connected services remain subject to their own charges.
From 1 December 2026, sandbox compute is billed at USD 0.0994 per vCPU-hour and memory at USD 0.0118 per GiB-hour. Microsoft says prices may vary by region and agreement.
Availability also depends on supported regions, models, tools and preview limits. Teams planning pilots should check those limits against the Foundry project, model deployment and connected services they intend to use.
Choosing the right agent model
The release gives developers three broad paths in Foundry Agent Service. Prompt agents without the Copilot harness suit tasks where instructions and supported tools are enough. Prompt agents with the Copilot harness add skills, managed execution, files, subagents and more tool coordination without requiring the team to own the harness.
Hosted agents are for cases where the team needs to bring or change the framework, agent loop, middleware, dependencies or runtime behaviour. The cost and engineering trade-off is different in each case, especially when the workflow has long running time, external writes or review requirements.
Choosing between a managed Copilot harness, a hosted agent loop and explicit recovery controls is the kind of agentic AI design decision covered in ECS's Core AI topic, alongside Azure AI Foundry, the EU AI Act and responsible AI. Sessions and workshops give attendees room to compare model choice, tool permissions, evaluation practice and practitioner experience with agents. Bring the agent patterns and failure cases that are difficult to settle from documentation alone.
This article was written by the ECS editorial team with the assistance of Amal Agent, run.events' AI assistant, which supported the research, drafting and proofreading.