A managed preview, with separate release stages
Microsoft's October 28 announcement introduces the Copilot harness in Foundry Agent Service as a public preview. Microsoft operates the agent loop and execution environment; developers supply instructions, skills, and tools. This is a managed route for prompt agents, rather than a new standalone language model. [1][3]
Keep the release stages separate. Long-running resilience for hosted agents is generally available; the Copilot harness preview is a different option. The broader Copilot SDK's capabilities do not establish what this managed service supports. [1]
The promotion covers runtime, not the whole task
Microsoft says harness runtime usage is free before December 1, 2026. Model inference, separately billed tools, and connected services still carry their applicable charges. Consumption-based runtime billing begins December 1, with prices varying by region and agreement. [1]
Access also depends on a supported region and a reasoning-and-tool-calling model available to the project. Bring your own model is unsupported. The setup guide requires a new prompt agent; existing agents cannot switch to this harness during the initial preview. [2][3]
Check what the conversation stores
Requests require a conversation with stored responses: store:false and previous_response_id are unsupported. Foundry Memory is unavailable, which does not mean conversation history or workspace files disappear. The preview documentation provides no service-level agreement and advises against production workloads. [2][3]
Our take
For experienced technicians, a managed loop can reduce infrastructure work while leaving important decisions with the team. Check the actual project region, stored data, tool permissions, and complete task cost before a small pilot. A runtime promotion does not make external actions free or appropriate.
We reviewed public sources; we have not deployed or tested this service.