Microsoft Copilot Harness Opens to Enterprise Developers

30-Second Summary

Microsoft opened the agent infrastructure behind its Copilot products to enterprise developers on 28 October 2026. The Copilot harness, which runs Copilot Studio, GitHub Copilot, the VS Code extension, and more, is now available as a public preview in Foundry Agent Service at no charge through 30 November 2026. Developers can build agents that surface in Microsoft 365 and Teams without rebuilding for each platform. Billing starts 1 December.

At a Glance

• Topic: Agent Systems
• Company: Microsoft
• Date: 28 October 2026
• Announcement: Copilot harness enters public preview in Foundry Agent Service
• What Changed: The orchestration layer behind Microsoft's own AI products is now available to enterprise developers for the first time
• Why It Matters: Any organisation on Microsoft 365 can now build agents with the same infrastructure that powers Microsoft's own products and deploy them to Teams without leaving the Microsoft stack
• Who Should Care: Technology leads, operations managers, and IT decision-makers at organisations running Microsoft 365 Business or Enterprise

Key Facts

• Microsoft announced the Copilot harness in Foundry Agent Service on 28 October 2026 (Source: Microsoft Official Blog, 28 October 2026)
• The harness is in public preview at no charge through 30 November 2026
• Billing starts 1 December 2026 at $0.0994 per vCPU-hour for sandbox compute and $0.0118 per GiB-hour for memory, with model inference charged separately
• The harness powers Copilot Code, Copilot Studio, GitHub Copilot app, the CLI, and VS Code
• Over 20 frontier models are supported
• Agent-to-agent (A2A) support is in public preview
• Agents can be published from Foundry to Microsoft 365 Copilot and Teams without being rebuilt
• Foundry Memory is not available during the preview period

What Happened

Microsoft announced the Copilot harness on 28 October 2026 in a post on the official Microsoft Blog. The harness is the orchestration layer the company uses internally to run Copilot Studio, GitHub Copilot, the VS Code extension, Copilot Code, and the CLI. That same infrastructure is now available to enterprise developers via Foundry Agent Service in public preview.

The announcement positions Foundry as the developer-led path and Copilot Studio as the business-led path. Developers working in Foundry supply their own domain knowledge, instructions, and tools. Microsoft manages the execution loop and sandbox. Agents built this way can be published to Microsoft 365 Copilot and Teams without being rebuilt for each surface.

Agent-to-agent (A2A) support also entered public preview with the announcement. A2A allows developers to build multiple specialised agents and wire them together, with one agent handing off to another based on the task at hand. The harness handles the routing, so developers do not need to build that orchestration from scratch.

The pricing window is the detail most likely to drive near-term action. The harness runtime is free through 30 November 2026. Billing starts 1 December at $0.0994 per vCPU-hour for compute and $0.0118 per GiB-hour for memory. Model inference is charged on top of that.

Why It Matters

The agent layer is becoming infrastructure. Microsoft is not building consumer features. It is opening the runtime that runs its own products to developers. This is the same strategic move Amazon made with AWS: take internal infrastructure, standardise it, and sell access to it. The Copilot harness becoming a developer product means agent orchestration is shifting from something you build to something you buy.

Distribution matters as much as capability. Most enterprise AI projects fail because employees do not change their behaviour. An agent that surfaces inside Teams and Microsoft 365, where employees already spend their day, removes the adoption friction. Developers no longer need to build a custom interface and convince staff to open it.

Multi-model support changes the procurement picture. Support for 20-plus frontier models means organisations are not locked into a single vendor's pricing or capability curve. A team that chose Claude for specific tasks can continue using it inside the Foundry infrastructure without switching runtimes.

A2A is the step toward autonomous workflows. The agent-to-agent preview is the most forward-looking element of the announcement. It points toward workflows where a lead qualification agent passes a prospect to a research agent, which passes a summary to a drafting agent, without a human initiating each handoff.

The free window is an experiment budget. Four weeks of free compute is not long enough to build production infrastructure. It is long enough to run one proof of concept, understand real costs, and decide whether to invest. The window is worth taking deliberately, not waiting out.

Small teams benefit from shared infrastructure. A 20-person firm cannot build and maintain a custom agent orchestration layer. It can build an agent definition and let Microsoft maintain the runtime. That asymmetry is what the harness opens up.

The David and Goliath View

The pattern here is familiar: a hyperscaler opens internal infrastructure to the market, lowers the build cost for everyone, and shifts competition to the application layer. The implication for small and mid-size businesses is that the right question is not whether you can build this. It is what you should build first, and what you need in place before 1 December to know whether the spend is justified.

For most businesses on Microsoft 365, the answer starts with one workflow. Not a strategy, not a roadmap, one specific task where an agent living in Teams would save a measurable amount of time. The free window is there to find that task, not to plan around it.

The A2A preview is worth watching even if you have no immediate use for it. It is Microsoft signalling the architecture it expects enterprises to build toward: not monolithic AI assistants, but coordinated teams of specialised agents. That is the architecture David and Goliath's Employee Amplification Systems work is built on, and the underlying infrastructure is converging.

Where This Fits in the AI Stack

The Copilot harness sits at the agent orchestration layer, one level above the model and one level below the business application. It handles the execution loop, the sandbox, tool connections via MCP, and the routing between agents. Developers supply the instructions, domain knowledge, and tools. Microsoft manages the runtime.

This is distinct from raw API access, where you manage every component yourself, and from Copilot Studio, where a business user builds flows without code. Foundry is for teams who can write and maintain agent definitions but do not want to build orchestration infrastructure from scratch.

Questions Operators Are Asking

Is this relevant if we already use Copilot Studio? Not immediately. Copilot Studio is the right starting point for business-led teams without an internal developer. Foundry is a separate product aimed at developer teams. If you are already getting value from Copilot Studio, the Foundry announcement does not change your workflow. If you have a developer and want to build more complex, multi-agent workflows, Foundry is worth evaluating before 1 December.

What does it cost after November? Compute is charged at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour starting 1 December 2026. Model inference costs are charged separately on top of those rates. The actual spend depends on your agent's workload. The free period is specifically designed to let you measure usage before billing begins.

Can we use models other than Microsoft's? Over 20 frontier models are supported. The harness is designed to be model-agnostic, so teams already committed to a particular model, including Claude, Gemini, or others, can continue using it. Note that bring-your-own-key support was flagged as potentially inconsistent in early reports. Confirm in Microsoft's current documentation before building around it.

How does A2A work in practice? Agent-to-agent support in public preview allows one agent to hand off a task to a specialised agent. An intake agent can collect context, delegate to a research agent, and pass the result to a drafting agent. The harness handles the routing. This is currently in preview, so the feature set and reliability are still stabilising.

What is not available yet? Foundry Memory is not available during the preview period. If your use case requires persistent memory across sessions, you will need to manage that layer separately or wait for it to be enabled.

Citable Summary

Microsoft opened its Copilot harness to enterprise developers via Foundry Agent Service on 28 October 2026. The harness is the same orchestration infrastructure behind Copilot Studio, GitHub Copilot, and VS Code. It is in public preview at no charge through 30 November 2026, with billing starting 1 December at $0.0994 per vCPU-hour plus memory and model inference costs. Agents built in Foundry can be published to Microsoft 365 and Teams without being rebuilt. Agent-to-agent (A2A) support is also in public preview. Foundry is code-first and aimed at developer teams; business-led teams should start with Copilot Studio. Sources: Microsoft Official Blog, 28 October 2026 (blogs.microsoft.com); Foundry Developer Blog, 28 October 2026 (devblogs.microsoft.com/foundry).