Products#MCP#Coding agent#Runtime
Microsoft opens its GitHub-built Copilot harness via Foundry: 20+ frontier models, self-hostable SDK
Microsoft opens the Copilot agent layer via Foundry: 20+ frontier models, BYOK, an MIT SDK to self-host, billing from December 1.
On October 28, 2026, Microsoft opened up the agent orchestration layer behind its Copilot products, the Microsoft Copilot harness, through Foundry. The official blog defines the harness as everything around the model — “if the model supplies the intelligence, the harness is everything around the model” — covering instructions, context, execution environment and controls. The announcement claims 20+ frontier models with bring-your-own-key support, an SDK open-sourced under MIT at github/copilot-sdk (about 10.6k stars as of October 29), and managed execution in Foundry that starts billing on December 1.
Key points
- Microsoft announced on October 28 that the same harness driving Copilot Code, Copilot Studio, the GitHub Copilot app and the CLI is available through Foundry, with a claimed 20+ frontier models, BYOK and per-task model routing.
- The copilot-sdk is MIT-licensed and ships for TypeScript, Python, Go, .NET, Rust and Java; it had about 10.6k GitHub stars as of 2026-10-29, and its README says it is becoming the Microsoft Copilot SDK and moving to the Microsoft org.
- Per Unite.AI, Foundry’s managed execution is free until December 1, 2026, then bills $0.0994 per vCPU-hour plus $0.0118 per GiB-hour; during the preview, agents cannot switch harnesses after creation.
Background
The harness is the engineering ring around a model in an agent runtime: how context gets fed in, where tools execute, how long a task runs, what happens on failure. GitHub’s team built this one and runs it in production across the Copilot family; Microsoft’s blog calls it “battle-tested and proven performant at the scale of millions of users.” The harness layer has been getting serious attention across the industry — we covered Pi taking its minimal terminal agent harness to 1.0 earlier this month, and Anthropic and OpenAI keep tightening the same layer inside their own coding agents.
The opening had a runway. On September 4, GitHub showed HydraFusion, a multi-model orchestration research preview combining models in single, cascade and critique patterns. On September 16 it disclosed the runtime port to Rust. Then on October 28, GitHub Universe’s official blog shipped the same capability to users under the name Copilot Fusion, the same day as Microsoft’s open announcement.
The facts
Microsoft’s one-line definition: the harness supplies instructions, context, an execution environment and controls around the model. Through Foundry, Microsoft runs the agent loop and the sandbox with governance built in. With the SDK, the same loop runs on your own machine.
| Item | Detail |
|---|---|
| Preview status | Public preview in Foundry Agent Service, from 2026-10-28 |
| Models | 20+ frontier models claimed, per-task routing; no model list published; managed Copilot harness does not support BYOM |
| Managed billing | Free until 2026-12-01; then $0.0994/vCPU-hour + $0.0118/GiB-hour (per Unite.AI, varies by region and agreement) |
| SDK | MIT license, six languages; Node.js, Python and .NET bundle the Copilot CLI automatically |
| Preview limits | Harness locked at agent creation; authoring via azd, the Foundry portal, the VS Code toolkit or the Copilot app |
The Rust port has numbers attached: per Microsoft’s September 16 disclosure, the runtime reached 832,378 lines of production Rust by August 21, shipped across 128 incremental PRs. Benchmarks run through the C# SDK put session creation plus one turn at 292 milliseconds, down from 5.25 seconds, and peak added memory in a ten-client batch fell from 1,383 MB to 126 MB. A Microsoft Distinguished Engineer ran those benchmarks — vendor-reported figures, in other words.
How to try it
The fastest path is the SDK: npm install @github/copilot-sdk (same idea for Python and .NET — the CLI comes bundled; Go, Java and Rust need the Copilot CLI installed separately). Signed in with GitHub, you need a Copilot subscription; configure BYOK and you can skip the subscription entirely, pointing the SDK at your own OpenAI, Anthropic or Microsoft Foundry keys. The README notes BYOK is key-based only — no Entra ID, no managed identities. For managed execution, create agents during the preview through the Azure Developer CLI (azd), the Foundry portal or the Foundry Toolkit for VS Code.
What others say
GitHub’s own release moves faster. The Universe blog, published the same day, brings third-party benchmark numbers for Copilot Fusion: on the Artificial Analysis Coding Agent Index, Fusion Max scored within two points of Opus 5.5 Max at half the cost and 40% less time per task, and Copilot code review Max posted the highest F1 score on ReviewBench for high-severity issues. HydraFusion’s September research preview reported +4.9 percentage points of verified task quality on TerminalBench 2.1 at an estimated 67% lower cost than Claude Opus 5. All vendor-selected benchmarks — read them accordingly.
Unite.AI is direct about the preview’s limits: the harness is chosen at creation and can’t be swapped; conversation history alone won’t restore in-progress work, so restarts need duplicate-action safeguards of your own. The New Stack adds a distinct model constraint: for the managed Copilot harness, the harness is selected when a prompt agent is created and cannot be changed later, even in a new version; developers may choose compatible models deployed through Foundry, but cannot bring their own models (BYOM). This is separate from BYOK support in the self-hosted SDK path. Microsoft’s framing is openness — the company says its platforms stay open to other harnesses and frameworks, with MCP, agent skills and plugins as open standards.
Our take
For teams that don’t want to write their own agent loop, this is a ready-made foundation: orchestration, sandboxing and model routing maintained by someone else, with BYOK as the escape hatch from Microsoft’s billing. The cost depends on the exit you pick. Self-hosting the SDK means owning the operations; handing execution to Foundry means metered vCPU and memory charges from December 1, plus a harness you can’t change once the agent exists.
Two things deserve skepticism. Microsoft’s blog never names the 20+ models, so “20+” remains a vendor claim for now. And the naming is muddled: the same capability is HydraFusion in links from Microsoft’s blog and Copilot Fusion in GitHub’s official Universe post, with no official explanation of the relationship. On the execution-environment front, Alibaba Cloud’s Agent Sandbox disclosure last month attacked a different layer — making sandboxes themselves a public utility at 100k per minute with sub-180ms P99 cold starts. Orchestration and sandboxing are both competing to be the agent infrastructure you build on; compare both before committing. As for what multi-model routing actually buys, Claude Sonnet 5.5 beating Opus 5.5 at release showed a single current model often suffices — weigh the routing savings against the complexity of calling multiple models.