Microsoft Copilot Harness arrives in Foundry: what is preview, what is GA, and what will it cost?

Microsoft's October 28 Foundry announcement describes an execution harness surrounding the model. It coordinates context, agent instructions, skills, tools, a sandbox and managed task loops. Developers can select supported models without rewriting all of this orchestration. The news is about the operating layer around an AI model, not a claim that Microsoft invented another foundation model. For AgentenCode, this makes the control, resilience and cost of the harness a more useful comparison than raw model benchmarks alone.

Microsoft lists reusable skills, connected tools including supported Model Context Protocol integrations, execution with bounded delegation to subagents, and source-controlled agent definitions. Microsoft runs the agent loop and sandbox while developers provide domain knowledge, configured credentials, permissions and tests. Public preview does not mean every model, region or integration works automatically. Tool approval and external account controls remain part of each deployment's design.

The original announcement contains two distinct releases. The Copilot Harness in Foundry Agent Service is in public preview. By contrast, long-running resilience for hosted agents is generally available. Hosted agents can use custom frameworks and execution loops. Foundry can restart interrupted work using the original input, but the application must persist checkpoints and prevent duplicate side effects. A resumed workflow can still send a message twice or create a duplicate record unless its own code checks what has already happened.

Microsoft says billing for harness runtime begins December 1, 2026. Runtime is free before then, but model inference, paid tools and connected services can already incur charges. The published December reference rates are USD $0.0994 per vCPU-hour and $0.0118 per GiB-hour for sandbox compute and memory, subject to region and agreements. These rates are resource units, not an all-inclusive agent price. Workload duration, parallelism and model/tool calls will affect a practical budget.

AgentenCode compares four choices: simple prompt agents, the new managed harness, hosted agents with custom runtimes and self-managed infrastructure. The right choice depends on complexity, operational ownership and measurable costs. Teams should calculate how many tasks run concurrently, how often they restart, how long tools execute and what data they need. Vendor marketing about multi-model support does not demonstrate that migration is effortless or that every workload will cost less than a custom solution.

The harness supports connected tools, but that does not grant permission to write into production systems by default. Identities, credentials, approvals, logging and data access must be explicitly configured. Long-running jobs introduce failure modes around stale context, retries and duplicate external actions. Microsoft provides recovery infrastructure; the operator still owns the business-level correctness of a process and the side effects it causes. This is why an isolated trial should include interrupted and resumed tasks, not just a happy-path demonstration.

The preview is relevant to development teams already building multi-step agent workflows with skills and tools, provided they can test within a controlled environment. Teams doing simple question-answer tasks may not benefit enough to justify a new managed layer. A meaningful pilot should include task outcomes, permissions, retry behavior, total model/tool spend and sandbox resources. We have not conducted such a benchmark ourselves. For a related permission analysis, see our GitHub Copilot Automations report.