SIGNAL / TOOLS

Microsoft Copilot Harness in Foundry: how it works and what costs

SOURCEMicrosoft

PUBLISHEDOctober 28, 2026

LAYERTools

SIGNAL SCORE70/100

SXF VERIFIED BRIEF

Microsoft brought its GitHub-built, multi-model Copilot agent harness to Foundry Agent Service in public preview. Instead of operating a separate agent execution loop, developers can define instructions, reusable skills and connected tools while Microsoft manages orchestration and sandbox execution. The Foundry announcement separately says long-running resilience for hosted agents is generally available; that does not make the harness itself generally available.

WHY IT MATTERS

The practical question for AI agent teams is what to build versus buy: a managed harness can reduce the engineering needed for tool calls, multi-step workflows and execution environments. It does not remove the need to evaluate supported models, tool permissions, data governance, task success and operating cost. This update gives teams a new deployment option, not proof that every workload gets cheaper or more accurate.

WHAT TO VERIFY

Check whether your Foundry region and model support the harness, whether your application needs preview features or a custom hosted agent, and how tool approvals and MCP credentials work. Microsoft says the harness is chosen when a prompt agent is created and cannot yet be switched on an existing agent. Runtime billing starts December 1, 2026, while separate model and tool charges can apply earlier.

SXF RESEARCH EXPLAINER

What the technical evidence means.

Source-grounded analysis

TECHNICAL ANALYSIS / 01

What a managed agent harness actually replaces

An AI model produces individual responses; a harness handles the surrounding system needed to finish a task. That includes keeping state between tool calls, giving the model the right working context, executing tools, moving files, and deciding what to do next. Microsoft's Foundry preview makes these capabilities available as a managed runtime instead of requiring each developer team to implement and maintain the same orchestration code. The underlying prompt agent remains the deployment object rather than becoming a new model.

This distinction matters when comparing a simple chatbot with a multi-step coding or research agent. Reusable skills package task procedures; MCP tool connections supply external capabilities; execution and delegation let an agent perform steps in a controlled environment. The actual business value depends on whether those parts complete an organization's own tasks reliably, not merely whether the agent can call more tools.

Check primary-source evidence ↗

TECHNICAL ANALYSIS / 02

Managed harness versus hosted agents: a practical choice

Foundry's documentation draws an explicit boundary. A prompt agent without this harness works for straightforward instructions and supported tool use. A prompt agent with the harness is intended for multi-step execution involving tools, files, skills or commands without operating the harness yourself. A hosted agent remains the better fit if you must bring your own orchestration framework, middleware, runtime dependencies or custom process model.

The Copilot harness is a preview, not a blanket production recommendation. Its support varies by region, model and tool integration, and Microsoft says the selection happens when the prompt agent is created. Existing agents are unchanged and cannot simply be toggled onto this harness. Before migrating a workflow, compare evaluation success rate, tool permission boundaries, latency and recovery from failed or interrupted tasks under the exact configuration that will ship.

Check primary-source evidence ↗

TECHNICAL ANALYSIS / 03

Cost, security and adoption checklist

Microsoft says Copilot harness runtime billing in Foundry begins December 1, 2026. Its announced sandbox rates are USD $0.0994 per vCPU-hour and $0.0118 per GiB-hour, subject to region and agreement; model inference and separately billed tools are not free just because the harness runtime is temporarily free. Measure actual sandbox lifetime and memory allocation on a representative workload before estimating monthly spend.

Start with a narrow, permission-limited pilot. Confirm the tool and MCP approval requirements, test prompts that attempt to change execution scope, log model and tool usage, and define a human handoff for consequential actions. Compare the harness with your current stack on completion quality and total cost rather than treating a vendor's broad model support or internal performance claims as an independent benchmark.

Check primary-source evidence ↗

PRIMARY-SOURCE FACTS

What the evidence establishes.

Reviewed 2026-11-02

Availability and launch

Copilot harness in Foundry Agent Service entered public preview on October 28, 2026; hosted-agent long-running resilience was separately announced as generally available. Primary evidence ↗

Managed runtime boundary

Foundry manages the agent execution loop and sandbox; customers supply agent instructions, skills, tools and model choices. Primary evidence ↗

Preview restrictions

The harness is optional for prompt agents and preview features have no production SLA; supported regions, models and tools are constrained. Primary evidence ↗

Model and orchestration scope

Microsoft says its shared harness supports many model families, reusable skills, MCP-connected tools and execution/delegation workflows. Primary evidence ↗

Pricing and start date

Foundry blog lists sandbox rates from December 1, 2026: $0.0994 per vCPU-hour and $0.0118 per GiB-hour, with regional and agreement variation; inference and tools can cost extra. Primary evidence ↗

LIMITATIONS

Microsoft's preview announcements and product docs are first-party sources; they do not independently demonstrate better task completion or lower cost across real workloads. Preview support, eligible regions, model availability and pricing may change. The separate hosted-agent long-running resilience GA claim must not be confused with harness GA.

TOPICSGitHub Copilot AI Agents Coding AI

MODELSNo named model detected

Microsoft Copilot Harness in Foundry: how it works and what costs

What the technical evidence means.

What a managed agent harness actually replaces

Managed harness versus hosted agents: a practical choice

Cost, security and adoption checklist

What the evidence establishes.

Keep the context connected.

Transformers Model Integration Contracts

Introducing the Grid Protectors Program

Route and orchestrate models with GitHub Copilot Fusion

GitHub intelligence puts your code, context, and history to work

Build shared workflows for your team with Copilot Automations