Microsoft opens Copilot harness preview in Foundry for managed AI agent workflows

Microsoft introduced the Microsoft Copilot harness in public preview within Foundry Agent Service on October 28, 2026. The managed infrastructure connects an AI model to tools, files and instructions for complex, sustained tasks. For Israeli companies developing service, marketing or operations agents, it offers a way to transfer part of execution management to Microsoft while retaining responsibility for the business process.

The same publication separately announced general availability of long-running resilience for hosted agents. The distinction matters: the new Copilot infrastructure remains in preview, while recovery belongs to the hosted agents route. These are options within one platform, with different status and limitations. Microsoft states in its documentation that the preview has no service-level agreement and is not recommended for production workloads.

What the infrastructure around the model does

A model can suggest an answer or workflow. Completing a task may require reading files, using outside services, remembering completed steps and deciding how to continue. Microsoft calls the infrastructure organizing these activities a harness. It says it runs the agent loop and isolated execution environment, while customers supply instructions, procedures and tools.

The broader announcement describes infrastructure developed at GitHub and adopted across Microsoft products. In Foundry, the offer concerns developer-defined agents and supported models. It does not establish that every model or tool is available in every region, or that existing applications can migrate unchanged. Availability and limits require checking against the particular design.

The useful business question is what remains the customer's responsibility. Even with provider-operated infrastructure, businesses must define correct results, permitted information and actions requiring approval. Shared infrastructure may reduce repeated development work, but does not define refund policies, service conditions or brand voice.

Four main preview capabilities

Microsoft identifies four components: reusable skills, connected tools, execution and delegation, and agent definitions managed alongside application code. Skills combine instructions, scripts and supporting materials. Connections include supported tools and MCP, with explicitly configured authentication and approvals. Agents can work with files and code and delegate bounded tasks to subagents.

In business terms, the requested task can be separated from its procedure. In a hypothetical market briefing, instructions define the subject, the procedure establishes acceptable sources and tools provide access to information. That separation makes methods easier to update without rewriting every request.

It also supports clearer quality discussions. When an output is wrong, teams can examine whether a source was missing, the procedure was ambiguous or a tool returned unsuitable data. These are recommended diagnostic questions, not a promise that every fault is detected automatically. Organized components still need an overall result check.

An interrupted task need not start from zero

The separate long-running resilience capability applies to hosted agents. Microsoft says Foundry detects interruptions and restarts work with the original input. Customer code must save progress and intermediate results so the agent can resume appropriately. Conversation history alone, the company stresses, is insufficient to restore work in progress.

This matters to companies developing automation beyond a brief demonstration. Hypothetically, an agent might analyze a file and create a record in another system. After an interruption and restart, the workflow needs to know that the record exists. Otherwise, repeating the request could duplicate the action. Microsoft explicitly says workflows must prevent such duplicates.

Define checkpoints and a way to establish whether an action has already happened. Business users need not understand every implementation detail, but should know whether restarting is safe. Agent reliability includes behavior after interruption, beyond the first successful run.

Costs have several components

Microsoft says usage-based billing for the Copilot harness runtime in Foundry Agent Service begins December 1, 2026. Runtime usage before that date is free, while model inference, separately billed tools and connected services remain subject to applicable charges. Preview access therefore does not mean the entire operation has no cost.

The announcement lists base rates of USD 0.0994 per vCPU-hour and USD 0.0118 per GiB-hour of memory from the billing date. Prices may vary by region and agreement. These are announcement figures checked on November 5, not a shekel quotation or a price commitment for an Israeli project.

When estimating expenditure, request a cost per completed, reviewed task rather than only per model call. A task may involve several processing stages, tools, storage and checks. Retries and corrections consume resources too. Without defining the workflow, the overall economics of managed infrastructure are difficult to assess.

Who the managed route suits

Microsoft distinguishes agents that need only instructions and supported tools, agents using the Copilot harness, and hosted agents bringing their own execution framework. The difference concerns how much control and engineering work customers need or want. A more complex route does not guarantee better business results for a simple task.

During preview, Microsoft says the harness is selected when creating an agent; switching it afterward is not yet supported. Existing agents are unchanged. Microsoft also manages execution environment CPU, memory and idle settings. These constraints should inform any migration plan.

Israeli companies should begin with a bounded process and checkable information. A draft report or internal summary can provide a learning environment before customer-facing action. Later, an AI workflow demonstration can show expected errors and human handover as well as the successful path.

Define ownership of outputs in advance. Whoever requests a report should know who approves its figures; whoever receives a draft response should know whether it may be sent. Without that division, technical progress can still stall between teams. A clear handover establishes whether the agent shortened total work or merely shifted review effort to another employee.

Development gets shorter when the process is defined

Assess development work saved against work remaining: tools, permissions, testing, records and support. The reviewed sources do not justify promising a fixed saving for every Israeli business. Benefits depend on existing infrastructure and the complexity of the tasks being assigned.

Microsoft's move lets developers purchase more of their execution environment as a service. For business owners, it sharpens the discussion: who operates the agent, who checks the result and who is responsible when work stops? With those answers defined, the technology can be tested on a real task at an explainable cost.