# GitHub Universe: Copilot became the agent harness > Source: > Published: 2026-10-29 12:18:45+00:00 GitHub Universe: Copilot became the agent harness GitHub shipped Copilot Fusion, multi-repo agent projects, computer use, and local model routing inside one desktop app at Universe 2026 on October 28. The message from Jay Parikh’s keynote, delivered by a Microsoft exec who now runs GitHub inside the CoreAI organization, is that GitHub stops being a place where developers write code and becomes the orchestration layer where agents do. Microsoft’s own blog post the same day names the thing plainly: the harness. If the model supplies the intelligence, the harness is everything around the model, and Microsoft wants to own that layer across every Copilot surface, from the desktop app to VS Code to the CLI to Foundry. ## What shipped, in one list The keynote post bundled a lot, and the changelog carried the details the same day: - **Copilot Fusion, generally available.** Pick Fusion instead of a model and GitHub chooses models per prompt across three effort levels (Standard, High, Max) and three execution patterns: Single, one model solves it; Cascade, a cheap model drafts and a judge accepts or escalates; Critique, one model drafts, a second reviews read-only, the first revises. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max at about half the average API cost per task. Billing is per model used at standard rates, with a limited 30 percent discount for Fusion traffic. - **Copilot projects, public preview.** Describe a goal, attach repos, design mocks, Slack threads. Copilot spins up as many agents as the work needs, runs them in parallel managed sessions, and hands back pull requests. The demo project ran nine PRs at once across multiple repos. - **Computer use, public preview on macOS and Windows.** Copilot can read app content, click controls, and type into desktop apps that have no API, no CLI, and no MCP integration. Every takeover requires approval, and org admins can disable the feature entirely. - **The GitHub Copilot app, out of technical preview.** Announced June 2 at Microsoft Build as a bare terminal window, now a full desktop client: a My Work view of active sessions, issues, PRs, and automations, chat surfaces, and canvas views where an agent’s plan or a browser session is a work surface you edit, not a chat scroll you squint at. - **Copilot code review Max.** A deeper review tier that GitHub says posted the highest F1 score on ReviewBench for high-severity issues, alongside the existing Lite and Balanced tiers. - **Enterprise bundling.** GitHub Enterprise now includes Copilot for every licensed user, replacing the separate Copilot subscription; enterprises pay for AI credits they use. The changelog notes it is available through GitHub Sales today, self-serve soon. ## The harness is the product The Microsoft blog post by Tina Schuchman and GitHub CTO Vladimir Fedorov explains why GitHub reorganized its whole stack around this. A frontier model today might fall behind tomorrow, and building a router to keep up is slow. GitHub’s answer is a harness layer built by the GitHub engineering team, now standardized across Copilot Code, Copilot Studio, the desktop app, the CLI, and VS Code. It supports MCP, agent skills, and plugins. It has been ported to Rust for lower session overhead and faster time to first token. It carries 20+ frontier models plus bring your own key. And the whole thing is now available through Microsoft Foundry, so enterprises can build their own agents on the same runtime. That is the strategic read: Microsoft stopped competing on “which model is smartest” and started selling the layer that makes any model useful. The keynote’s own framing says the quiet part: developers shouldn’t have to manage every model choice to get the right result at the right price. In practice that means OpenAI, Anthropic, and xAI models become interchangeable parts inside a GitHub-owned harness, selected per prompt by GitHub’s routing layer. The moat moves up the stack. ## The part this site cares about: local models ride the same harness Buried in the Fusion changelog is the announcement with the sharpest edge for local inference. Copilot Fusion now supports intelligent routing to on-device models, in experimental preview, in the Copilot CLI, the Copilot app, and VS Code. Two paths: - **Managed (Windows only):** Windows ML downloads and configures MAI Code 1.1 Flash on the device. The screenshot GitHub shipped shows the GGUF (MAI-Code-1.1-Flash-Q2-XS.gguf) downloading, 1.0 GB into 53 GB budgeted. - **Bring your own local model (Windows, macOS, Linux):** point Fusion at any OpenAI-compatible local server, configure MAI Code 1.1 Flash or Qwen3.8 Flash Next, and let Fusion choose between local and cloud per request. The Copilot app’s provider settings show a Custom endpoint provider type that explicitly names vLLM and OpenRouter as examples. Read that twice. The world’s default coding platform now treats a locally hosted model as a routing candidate alongside Claude and GPT, chosen per request by efficiency logic. The caveat matters and GitHub states it: local routing is not offline mode, Fusion still makes routing decisions in the cloud, and local models may have different safety protections from Copilot-provided ones. The direction of travel is the news: harness parity for local weights, not a walled garden with a local escape hatch. ## The numbers behind the pivot Octoverse landed the same day with the usage data that explains the strategy. Developers pushed code to GitHub 17.9 billion times in the past 12 months, more than 2024 and 2025 combined. Repos with an agent configuration file averaged eight times as many commits as repos without one. More than 40 percent of pull requests now start with an agent, up from 17 percent in March. And 61 percent of Copilot users touched more than one model in September. Multi-model is not a feature somebody requested; it is observed behavior GitHub is productizing. GitHub also put its infrastructure bill where the agent future is: 4 million CPU cores, 160 PB of high speed storage, and 50 Tb/s of network capacity added this year, plus a multi-region Azure migration expected to finish early next year. ## What it means if you run local models Three takeaways worth acting on. First, the harness layer is consolidating fast, and the two biggest platforms (Microsoft’s Copilot stack and Anthropic’s Claude Code) now both treat models as swappable parts, which means model choice matters less inside them than it did this time last year. Quality differences between frontier models are increasingly mediated by routing logic instead of user choice. Second, local models just got a distribution channel no open weights project could have bought. Qwen3.8 Flash Next listed as a first-class local routing target in Copilot’s own client is the kind of placement open source labs have been fighting for, and it is opt-in for every GitHub user with the compute. Third, the pricing change is underrated. Folding Copilot into Enterprise for all licensed users at credit-based billing means every GitHub seat becomes an agent seat by default. The audience for local model routing is suddenly everyone with a GitHub login, not the enthusiasts who configure llama.cpp by hand. The June technical preview shipped as a terminal window with tabs for Session, Issues, Pull requests, and Gists, three months before the polished desktop client. The pace of this product line is the tell: GitHub went from completions vendor to harness vendor in under a year, and the infrastructure numbers say they expect agents, not humans, to be the primary traffic.