What happened
Microsoft announced the Microsoft Copilot harness in Foundry Agent Service on October 28, 2026, as a public preview. The company describes it as a managed execution system for agents defined declaratively in Foundry Agent Service: Microsoft runs the agent loop and sandbox, while the developer supplies the knowledge, procedures and tools specific to an application. The system maintains working context as tasks develop and coordinates model and tool steps toward a requested outcome, within configured limits and policy boundaries. Developers can choose a supported Foundry model without rewriting the orchestration around it.
The preview combines four building blocks. Reusable skills package instructions, scripts and supporting resources into procedures an agent can apply across tasks. Connected tools support Foundry and Model Context Protocol (MCP) integrations, with authentication and approvals configured explicitly. Execution and delegation let agents work with files and code, and delegate bounded tasks to built-in subagents while a main agent coordinates the result. Source-controlled definitions let teams create and test in Foundry or manage agent definitions and supporting resources alongside application code.
Microsoft illustrates the authoring flow with a research-brief agent built using the preview Azure Developer CLI surface. A skill describes a research procedure; an azure.yaml definition specifies a prompt agent, the microsoft_copilot_preview harness, a model deployment and web-search and code-interpreter tools. Developers can also start in the Foundry portal or Foundry Toolkit for Visual Studio Code. In the GitHub Copilot app, a Foundry plugin and skill offer a conversational path to managing prompt agents. Microsoft says hosted agents remain an alternative when an application needs its own framework, agent loop, middleware, dependencies or runtime behavior.
The announcement also makes long-running resilience for hosted agents generally available. Foundry detects interruptions and restarts work with its original input; the agent's code must save progress and intermediate results to resume. Workflows that update other systems need to guard against duplicate actions. The preview harness runtime is free before December 1, 2026. After that, sandbox compute is priced at $0.0994 per vCPU-hour and $0.0118 per GiB-hour; model inference and separately billed tools cost extra. Existing agents are unchanged, and the harness cannot yet be switched on after agent creation.
What the archive adds
This is a new product event on a known runtime theme, not a restatement of one earlier release. Unreal Agent: Async-first agent harness treats persistent, long-running work as a runtime problem; Microsoft now presents a managed option for that layer. A previous Telegram post described Foundry Agent Service as compute for agent workloads. The Microsoft Copilot SDK transition is related at the company-platform level, but this announcement concerns a separate Foundry service.
Trend signal
This supports harness architecture outlives model choice: Microsoft explicitly separates its managed execution foundation from a developer's choice among supported Foundry models. It is evidence of a product direction, not proof that this harness is more durable or effective than alternatives. The skills become a portable capability layer pattern is also relevant: the preview packages procedures as reusable skills, though the announcement does not establish their portability beyond this environment.
Why now
Microsoft frames the preview as part of a unified harness strategy across its agent experiences and pairs it with general availability of resilience for hosted agents. The source gives no further explanation for the timing. Its practical distinction is between a managed harness for declaratively defined agents and hosted agents for teams bringing their own execution framework.
Connections to other events
A previous Telegram post described Foundry Agent Service as a compute environment for agents that write files, execute code and maintain context. This release makes the managed agent-loop layer more explicit. Separately, the GitHub Copilot SDK transition to Microsoft Copilot SDK reflects Microsoft's stated effort to broaden a shared agent foundation; the current announcement adds a Foundry-specific hosted harness rather than describing an SDK migration.
Why it matters
A managed harness changes the amount of agent infrastructure a team must own. Developers may be able to spend less time implementing orchestration, context handling and sandbox execution, and more time improving domain knowledge, procedures and evaluation. The preview's explicit tool authentication and approvals, configured limits and policy boundaries give teams concrete controls to examine when assessing whether a managed environment fits their workflows.
The trade-off is a platform-defined execution model. Teams should verify which supported models and tools meet their requirements, how the preview behaves under realistic workloads, and whether its constraints fit their security and operational needs. The post's customer testimonials describe early enthusiasm, but they do not provide comparable performance or cost evidence. Pricing also changes after December 1, 2026, and model inference and certain tools remain additional costs. Existing agents cannot simply enable the harness after creation, so adoption may require planning around agent definitions rather than a seamless toggle.
Limits and uncertainty
The announcement does not enumerate all supported models, tools, regions, quotas or preview limitations. It provides no independent benchmark, measured cost comparison, reliability rate or detailed security assessment. Customer quotations are testimonials, not controlled evidence of outcomes. The announced compute prices apply after December 1, 2026, while inference and some tools are extra; the material does not estimate a representative workload's total bill. Long-running recovery depends on agent code preserving progress, and the source warns that workflows updating other systems need duplicate-action safeguards. Microsoft also says existing agents are unchanged and the harness cannot yet be enabled after creation.
Skill to build now
Prototype one bounded workflow that runs longer than a single model call, such as an evidence-backed research brief. Separate the procedure into a reusable skill, connect only approved search and analysis tools, and specify what evidence and limitations the result must return. If testing hosted-agent recovery, persist progress and intermediate results, then verify that resumed work does not repeat external side effects.
Reader takeaway
Microsoft's Foundry preview packages agent orchestration, sandboxing, tools, skills and bounded delegation into a managed harness, while hosted agents gain generally available recovery from interruptions. Treat it as a new option to evaluate—not demonstrated evidence of better results. Check preview fit, recovery behavior and full costs, especially the additional inference and tool charges and the compute rates that begin after December 1, 2026.