Microsoft and GitHub unify their agents on a common orchestration layer
Microsoft opens Foundry, the orchestration layer that runs Copilot. GitHub also generalizes Copilot Fusion, its multi-model logic, with a new billing model and agent features closer to professional use.
On October 28, 2026, Microsoft made the Copilot orchestration layer available to developers and businesses via Microsoft Foundry, while GitHub released Copilot Fusion to general availability the same day. Behind these announcements, the group aims to standardize how its AI agents select models, execute tasks, and maintain access control. The goal is simple: to ensure the agent is no longer just a chatbot, but a controllable, measurable, and governable system.
Une même brique pour faire tourner les agents
Microsoft presents the Copilot harness as the layer surrounding the model: instructions, context, runtime environment, and safeguards. This component already powers Copilot Code, Copilot Studio, the GitHub Copilot application, Copilot CLI, and VS Code, according to the Official Microsoft Blog. The company also states that the system supports open standards like the Model Context Protocol, as well as agent skills and plugins.
In practice, the harness manages the agent loop and sandbox, with built-in governance and optimization. Microsoft specifies that over 20 cutting-edge models are compatible, with the option to bring your own key, and that the platform distributes tasks to the most suitable model(s). The Copilot SDK is downloadable, and the same engine can also run in Foundry, hosted by Microsoft.
GitHub Fusion pousse le multi-modèle en production
At GitHub, Copilot Fusion replaces the simple choice of a model with a routing and orchestration system. When a user selects Fusion in the selector, GitHub decides which models to use and how to make them work together. Three effort levels are offered: Standard for common tasks, High for more difficult cases, Max for the most complex requests.
The system incorporates three execution modes already tested with Project HydraFusion: Single, where one model handles the request alone; Cascade, where a lighter model prepares a response before possible escalation; Critique, where a second model reviews before revision. GitHub states that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, at about half the API cost per task compared to a competitor cited in its tests, with roughly half the tokens and 40% less time.
A subtlety: at GitHub, the setting no longer just chooses a model, it chooses a strategy. Less glamorous than a slogan, but much more useful when controlling costs at scale.
Ce que Foundry change pour les équipes
Microsoft also highlights an agent creation mode managed in Foundry Agent Service. The public preview version combines four blocks: reusable skills, connected tools, execution and delegation to sub-agents, as well as versioned agent definitions with application code. Developers can create these agents via Azure Developer CLI, the Foundry portal, the Foundry Toolkit for Visual Studio Code, or the GitHub Copilot plugin.
The service separates usages. Simple “prompt agents” suffice when instructions and tools do the job. Prompt agents with harness add managed execution, files, code, and sub-agents. Finally, “hosted agents” maintain their own framework and agent loop while remaining hosted by Foundry.
On October 28, 2026, Microsoft also announced general availability of long-duration resilience for hosted agents. The system restarts work after interruption from the original input, while client code must save progress. For workflows modifying other systems, duplicates must also be avoided on restart.
Prix, accès et ce que les clients vont réellement payer
Billing for the Copilot harness in Foundry Agent Service begins December 1, 2026, on a consumption basis. Before this date, harness runtime usage is free. Afterwards, sandbox compute is billed at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, with rates varying by region and commercial agreement. Inference and connected tool costs remain billed separately.
On GitHub’s side, Fusion is billed model by model at each model’s standard rate, with a temporary 30% discount for paying Copilot subscribers. The service is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, CLI, and JetBrains IDEs. Administrators can also restrict available models by organization policy without cutting off other account uses.
Pourquoi Microsoft standardise maintenant
Microsoft is not starting from scratch. GitHub has already ported the Copilot runtime from TypeScript to Rust, with 832,378 lines of production Rust as of August 21, 2026, and 128 PRs delivered progressively. The company attributes this port to significant gains in session overhead and time to first token. In tests cited by Stephen Toub, client creation, session, and turn time dropped from 5.25 seconds to 292 milliseconds, with maximum additional memory reduced from 1,383 MB to 126 MB over a batch of ten clients.
Meanwhile, GitHub added 4 million CPU cores, 160 PB of fast storage, and 50 Tb/s of network in 2026, according to its own figures. The strategy is clear: simplify primitives, separate storage and compute, reduce coordination. This also prepares migration to multi-region deployment on Azure, expected early 2027.
By the numbers
- 20+ models — compatible with the Copilot harness, according to Microsoft.
- $0.0994/vCPU-hour — sandbox compute billed from December 1, 2026.
- $0.0118/GiB-hour — sandbox memory billed from December 1, 2026.
- 30% — temporary discount on model costs for paying Copilot subscribers.
- 832,378 lines — Rust production base at GitHub as of August 21, 2026.