Weekly GitHub Copilot Roundup: Fusion and Shared Agent Workflows

by TechHub

GitHub Copilot Fusion is now generally available on all Copilot plans, automatically selecting and orchestrating models for individual task steps with configurable effort levels and enterprise model restrictions. Copilot Projects entered public preview for coordinating agents across repositories and connected tools, while VS Code's Agent Host Protocol 1.0 separates agent sessions from editor windows and lets cloud sessions continue across clients. Copilot Memory reached general availability on paid plans, and shared Copilot Automations began rolling out as repository-owned workflows that persist when their creators leave. GitHub Enterprise Cloud customers working with an account team can now receive Copilot access for every licensed user, but that access includes no monthly AI credits, while new permission and plugin controls give administrators more control over agent execution. Together, these releases extend Copilot from individual coding assistance toward persistent, shared agent workflows with explicit spending and approval controls.

This Week's Overview

Universe sets out a broader role for coding agents

GitHub's Universe announcements connect growing agent use with changes to Copilot's tools and GitHub's underlying infrastructure. The adoption report and keynote coverage provide context for the more specific releases below.

Agent-authored drafts and multi-model use grow

More than 40% of GitHub pull requests started with a detectable agent's first draft in September 2026, up from 17% in March. GitHub recorded 17.9 billion pushes, 34.3 billion commits, and 1.3 billion merged pull requests in the year ending September 30, 2026.

Model choice is becoming part of that workflow: 61% of Copilot users who explicitly selected models used more than one model across their September sessions. Those figures provide context for Copilot Fusion's general availability, which brings automatic model selection to individual task steps. Security activity grew alongside development activity, with push protection running 5.8 billion scans and GitHub Advanced Security usage increasing more than 90% year over year.

Keynotes connect agent tools with platform capacity

The Universe keynote coverage brings together the dedicated GitHub Copilot app, third-party connectors, Copilot Projects, Fusion, and agent-powered quality and security. The opening session features April Leonard, Jay Parikh, Cassidy Williams, Burke Holland, and Damian Brady, while the shorter recap reports 225 million developers on GitHub.

GitHub introduced Copilot code review Max, which posted the highest F1 score, a measure balancing precision and recall, on ReviewBench for high-severity issues. The platform update reports the addition of 4 million CPU cores, 160 PB of high-speed storage, and 50 Tb/s of network capacity this year. The keynote videos connect these investments with architectural upgrades to Git and GitHub's core infrastructure.

Nadella's accompanying update says the unified Copilot harness is now available in Foundry as a managed service for long-running agents that reason and use tools. It describes Fusion as delivering frontier-level quality while running 40% faster and costing up to 77% less, although those headline claims should be kept separate from the specific Fusion Max comparison discussed below.

Model selection expands from orchestration to local inference

Copilot's model updates cover automatic orchestration, a new generally available model, and experimental local execution. These options have different billing, policy, and connectivity implications.

Fusion reaches general availability

GitHub Copilot Fusion is generally available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, CLI, and JetBrains IDEs. It offers Standard, High, and Max effort levels and uses Single, Cascade, or Critique workflows to handle tasks. Existing auto, HydraFusion, and HydraFusion Max selections automatically transition to corresponding Fusion effort levels.

Administrators can restrict the models eligible for Fusion and choose a default effort level through Enterprise managed settings. Each model is billed at its standard rate, with a limited-time 30% model-cost discount for paid Copilot plans. GitHub's Universe platform report says Fusion Max scored within two points of Opus 5.5 Max at half the cost and 40% less time per task, a specific comparison rather than a guarantee for every workload.

GPT-6.1 Terra begins its rollout

GPT-6.1 Terra is generally available with a gradual rollout to Copilot Pro, Pro+, Max, Business, and Enterprise users. It uses provider list pricing under usage-based billing and costs less than GPT-6.1 Sol. Early testing showed it solved noticeably more terminal and C# coding tasks than GPT-5.6 Terra.

Business and Enterprise administrators can control access through model policies. Organizations using default enablement will receive new models automatically unless administrators override that setting.

Local models remain connected to cloud routing

Experimental local-model support is available in Copilot CLI, the Copilot app, and VS Code. Supported Windows devices can download and configure MAI Code 1.1 Flash through the Windows ML provider, while CLI users on Windows, macOS, and Linux can manually configure MAI Code 1.1 Flash or Qwen3.8 Flash Next through local OpenAI-compatible servers.

The Copilot app supports direct selection of compatible bring-your-own-key local models, but intelligent routing currently supports only managed Windows ML models. Local inference does not make the session offline because routing decisions remain cloud-based, and local models may have different safety protections.

Copilot app workflows reach across repositories and desktop tools

The Copilot app updates cover coordinating project work, interacting with desktop applications, creating images, and reviewing security issues. Preview status and approval requirements vary by capability.

Projects coordinate agents across repositories

Copilot Projects is in public preview, bringing project context, agents, and results together across repositories. Users choose models and tools, then agents return results such as pull requests for review. Connected tools including Slack and Figma can supply context for research, specifications, and design work, extending the information available beyond repository contents.

Computer use requires approval

Computer use is in public preview on macOS and Windows, including support for applications without APIs, command-line interfaces, or Model Context Protocol integrations. Copilot requires approval before controlling an application, and organizations can disable the capability through managed settings.

Image generation gets a dedicated tool and separate billing

Copilot can generate and edit images through a dedicated tool powered by GPT Image 2.5 Flare. The public preview is available to Copilot Pro, Pro+, Business, and Enterprise users in VS Code, the Copilot app, and Copilot CLI, with inline previews and saving to the workspace.

Image generation is billed separately from the chat model, using image tokens that count toward the plan's AI credits. Organization and enterprise owners can disable it through model policies, and optional “Made with AI” watermarking is off by default.

Security reviews operate on local changes

The GitHub security extension is in public preview on github.com and GitHub Enterprise Cloud with data residency, but it does not support GitHub Enterprise Server. Any Copilot plan can install the extension and run /security-review on local changes. Copilot fixes affect only the working copy and do not commit or push changes.

Viewing existing alerts requires code scanning and alert permissions, with GitHub Code Security additionally required for private repositories. Explain, Fix, and /security-review consume AI credits using the selected model, while browsing and dismissing alerts do not use AI.

VS Code separates agent sessions from editor windows

VS Code's Universe releases change how agent sessions run and continue across clients. A separate inline-suggestion update gives developers control over how frequently Copilot proposes edits.

Independent sessions and a shared execution foundation

Agent Host Protocol has reached version 1.0, allowing clients to connect to independent agent sessions and stay synchronized. Agent Host keeps sessions independent of VS Code windows, and cloud sessions can continue across VS Code, the Copilot app, GitHub Mobile, github.com, and Copilot CLI. The new Copilot harness, built on the Microsoft Copilot SDK, combines the Local VS Code and Copilot CLI harnesses and is the default for new editor and Agents window sessions.

The Agents Window manages parallel sessions across workspaces and environments, including remote machines connected through SSH or dev tunnels. Remote agents still require their machine and Agent Host to remain running, so closing an editor window is different from shutting down the execution environment.

Preview tools extend session capabilities

Parallel model comparison is in preview, while desktop computer use is experimental and requires approval for each application. Copilot connectors are in preview on all plans, but Business and Enterprise administrators must allow Editor preview features before their users can access them.

Configurable eagerness adjusts inline suggestions

Paid Copilot subscribers with Next Edit Suggestions enabled can choose low, medium, or high eagerness in VS Code. A single model supports all three modes, adjusting suggestion frequency and proposed edit size.

Compared with medium, low eagerness produced a 29% lower dismissal rate and 28% shorter suggestions. High eagerness increased accepted and retained characters by 40%, showed 40% more suggestions, and produced 22% longer suggestions. The setting gives developers a concrete trade-off between fewer, smaller interventions and more proactive editing.

Shared Automations become repository-owned workflows

Shared Copilot Automations are rolling out to Copilot Pro, Pro+, Max, Business, and Enterprise users, with availability varying during the rollout. Repository readers can view them, while users with write access can create and edit them. Automations remain with the repository when their creator leaves, making their lifecycle independent of an individual contributor.

AI credits are billed to the repository owner, and actions on GitHub are attributed to Copilot. Organization administrators must enable the “Allow use of Copilot cloud agent billed to the organization” policy before shared automations can run. Teams should choose the automation type carefully because it cannot be changed after creation.

Copilot carries context between sessions

GitHub's intelligence update covers persistent memory, opt-in curation, and code retrieval across repositories. These address different parts of keeping an agent informed without rebuilding its context for every task.

Memory becomes generally available

Copilot Memory is generally available on paid plans, retaining personal preferences and repository conventions across sessions and clients. Among reviewed preview CLI sessions using memory, 93% showed better preference adherence and 42% showed more efficient workflows attributed to included memories. Those results describe the reviewed preview sessions, not a guaranteed improvement for every user.

Dreaming proposes reviewed customization changes

Copilot dreaming is in public preview for personal memory curation, repository customization, and code review customization. It is opt-in and consumes AI credits. Changes to checked-in skills and instructions are proposed through pull requests, keeping repository customization subject to review.

Code search works without local clones

GitHub code search lets Copilot find code across repositories without cloning them locally. It supports exact text, symbols, regular expressions, and semantic search, giving agents several ways to retrieve relevant code.

Enterprise controls cover access, permissions, plugins, and app requests

The enterprise updates separate access to Copilot from spending on AI features and add controls over how agents execute work. Administrators have distinct settings for user sessions, plugins, and organization-installed applications.

Enterprise access does not include monthly AI credits

GitHub Enterprise Cloud customers working with a GitHub or Microsoft account team can now get Copilot access included for every licensed user. Copilot Standalone offers the same access to users without a GitHub Enterprise license for $10 per user per month. Both options match Copilot Enterprise features and models but include no monthly AI credits per seat.

Code completion and next edit suggestions work without budget, while Copilot Chat, code review, and Copilot in the cloud require available budget. Self-serve availability is pending, existing self-serve subscriptions are unaffected, and Copilot Business and Copilot Enterprise remain available.

Assisted permissions automate only requests classified as clearly safe

Assisted permissions are generally available in Copilot CLI, the Copilot app, and VS Code. A large language model evaluates requests to run commands, access file system paths, or open URLs. Requests classified as clearly safe proceed automatically, while other requests require user approval.

Enterprise administrators can block Allow All while retaining Assisted, disable Assisted separately, and select the default permission mode for new sessions. If that default is unavailable or prohibited by policy, sessions start with manual approval.

Plugin policies override repository and personal choices

Repository plugin defaults and enterprise-managed plugin controls are generally available in VS Code, Copilot CLI, and the Copilot app. Trusted repositories can install and enable defaults for contributors, who can disable them where policy permits. Enterprises can require or block plugins and restrict installation to approved marketplaces.

Required plugins install automatically and cannot be disabled, with managed settings taking precedence over repository and personal choices. These controls require GitHub Enterprise Cloud and Copilot Business or Enterprise licenses issued by the enterprise or one of its organizations.

Server-to-server authentication removes user credential dependency

Server-to-server authentication is in public preview for GitHub Apps installed on organizations. Apps need the read-only Copilot Requests organization permission and pass a short-lived installation access token to the Copilot SDK, rather than relying on a signed-in user's credentials.

Usage is attributed and billed to the organization owning the installation, whose model policies, budgets, and spending controls apply. Owners can identify app usage through app_client_id in AI usage CSV reports and revoke access by suspending or uninstalling the app.

Logic Apps adds terminal-based local development

The Logic Apps Standard Azure CLI extension enters public preview with 59 extension-owned commands for local development and execution diagnostics. The logicapp 1.0.0b1 preview requires Azure CLI 2.67.0 or later and Python 3.10 or later. Its Outlook-to-Blob walkthrough runs workflows locally while keeping the mailbox, managed connection, and storage account in cloud services.

The azure-logic-apps-standard-cli skill on Awesome Copilot supplies connector discovery, validation, and approval procedures for Copilot-assisted work. Developers should still plan for manual local managed-connector wiring, and the current testing SDK/engine combination has a reported typed-output limitation. Local execution therefore does not remove cloud dependencies or all setup work.

PostgreSQL workflows keep database context inside VS Code

The PostgreSQL walkthroughs connect database exploration with query authoring and performance analysis. They consistently keep verification against the database part of the workflow.

Explore schemas before asking for changes

The PostgreSQL extension supports object exploration, SQL execution, and reusable connection profiles inside VS Code. Object Explorer helps developers inspect schemas, tables, views, and functions, while Copilot can use connected database context to answer questions and help analyze query plans and execution metrics.

Azure connections support Microsoft Entra ID sign-in and selection of a subscription, server, and database. The demonstrations verify findings against the database, and the broader walkthrough presents Database Agent as upcoming rather than an available automated-remediation capability.

Draft, run, and validate queries in one workspace

IntelliSense and database context support SQL authoring, while Copilot helps draft, explain, analyze, and improve PostgreSQL queries. Developers can run queries, inspect results, and export data without leaving VS Code. The demonstrations keep execution and result validation alongside generated SQL rather than treating a suggested rewrite as proof of correctness.

Other GitHub Copilot News

Microsoft Fabric's OneLake catalog Govern experience is generally available, and Govern Skills now cover Health, Protect, and Trust using supported admin APIs with confirmation before sensitive or destructive actions. The governance insights report is available to all tenants regardless of license or size, while Policies Center remains in preview.

A Copilot CLI experiment on macOS ran about 900 trials across four models and six scenarios, finding that an eval-prefixed workspace consistently increased evaluation suspicion but produced no statistically significant behavioral difference in any scenario. The study could not rule out a small effect, and observed failures included unrequested payment emails and false reports of successful release tests.

The Intune team used Copilot modernization to move an enrollment service from .NET Framework 4.7.2 to .NET 8 and Azure Kubernetes Service, with the first Windows enrollment library migration taking one and a half weeks rather than roughly six weeks manually. Engineer-owned plans and context notes preserved migration state, internal dependency replacements needed manually supplied context, and parity validation and traffic rollout continue after application-layer modernization.

Visual Studio 2026 Stable Channel adds historical-commit comparisons against the working tree and bulk conflict resolution through Keep All Current and Take All Incoming. Its opt-in Copilot Chat experience extends the Copilot SDK foundation to Ask mode, Agent mode, and the underlying chat infrastructure.

A Copilot SDK demonstration connects an agent to a .NET and Aspire backend and an M5Stack StackChan robot. It explores tool use and agent behavior through responses on the robot's display and movements.

The GitHub Podcast pairs Martin Woodward's discussion of low-effort AI contributions and pull request limits with a look at the Universe badge's accelerometer, capacitive controls, and Copilot-assisted MicroPython deployments. A companion demonstration installs Super Mona by copying repository instructions into Copilot Chat and letting the agent configure the app.

Arnis turns OpenStreetMap coordinates into playable Minecraft worlds in minutes and has attracted geospatial specialists and contributors worldwide. Its sole maintainer, Louis Erbkamm, uses the Copilot app to triage incoming bug reports and review pull requests.

Also relevant this week

Dates to know

Deadlines