GitHub Copilot app: The agent-native desktop expertise

Whereas the agentic shift has made growth sooner, it’s additionally led to disjointed workflows, extra context switching, and an excessive amount of time spent reviewing agent-generated code.

If brokers are going to be a sturdy a part of how software program will get constructed, they want an actual place within the developer workflow. But most developer instruments weren’t designed for guiding a number of brokers in parallel. Context scatters throughout home windows. You lose monitor of what’s operating. Code lands in pull requests and not using a clear path of what the agent tried, what it validated, or the place human judgment is required.

Throughout GitHub, builders are utilizing brokers to maneuver from immediate to plan, from concern to drag request, from evaluate suggestions to merged code. As agentic workflows change into the norm, repository creation, pull request exercise, and API utilization are all accelerating with no proof of slowing down. On GitHub alone, commits practically doubled 12 months over 12 months, crossing 1.4 billion per 30 days, plus over 2 billion GitHub Actions minutes per week.

To satisfy this demand and proceed to be the house for all builders (and now their brokers), our focus is scaling our underlying programs and enhancing resilience and stability throughout all of our companies, at each layer of the stack.

GitHub is constructing that system for the agentic frontier, and that’s what we’re exhibiting as we speak at Microsoft Construct.

Copilot app: A management middle for agent-native growth

You begin the day with three items of labor already in movement. One agent is investigating a manufacturing bug. One other is implementing a backlog concern. A 3rd is working by way of evaluate suggestions on a pull request. Every is operating in its personal remoted setting, producing modifications you possibly can examine, redirect, check, and merge.

You want an setting that may sustain.

The brand new GitHub Copilot app is the agent-native desktop expertise constructed on GitHub. From a single My Work view, you possibly can see work in movement throughout related repositories: lively periods, points, pull requests, and background automations. The Copilot app is now obtainable in technical preview for current Copilot Professional, Professional+, Enterprise, and Enterprise customers.

The GitHub Copilot app is the newest in a line of AI tooling from GitHub that’s reworking our enterprise. Shifting past AI help, the app has supplied a much-needed management middle for agentic growth.

Our Ahead Deployed Engineers can dispatch a cohort of brokers and handle a number of initiatives, all from one location. Quick access to plans and autopilots with the power to run interactive periods or step into code the place wanted.

David Jobling | Grasp Expertise Architect, Head of Expertise & Supply Futures, International Solutioning & Supply, Avanade Inc.

Each session runs in its personal git worktree, an actual, remoted copy of your department. This helps parallel agent periods work with out stepping on one another. The app handles each worktree for you: no handbook setup, no cleanup, no department juggling. Whether or not you begin from a immediate or a problem out of your inbox, Copilot will get the context it wants from current points, pull requests, and the repos you’ve related.

Improve your Copilot subscription

For agent energy customers, Copilot Max is accessible as an improve from Copilot Professional, Professional+, and EDU, and is designed to help higher-volume utilization with out interrupting that circulation.

Study extra >

Then Agent Merge helps carry that pull request by way of evaluate, checks, and merge. It displays CI, tracks required reviewers, addresses failing checks, and waits for all situations to be glad. You select how far Copilot ought to go: drive CI again to inexperienced, handle suggestions, or merge when your situations are met. You determine what automation is enabled and what ships.

Canvas: The place intent turns into inspectable work

Chat is highly effective for instruction and ambiguity. However as soon as an agent begins doing actual work, a chat thread turns into an extended scroll of choices, logs, and corrections. You want a spot the place the work itself is seen.

As we speak, we’re additionally introducing canvases within the GitHub Copilot app. Canvases are bidirectional work surfaces for people and brokers. A canvas would possibly present a plan, pull request, browser session, terminal, deployment, dashboard, or workflow state. Brokers replace the canvas as they work, and builders can edit, reorder, approve, or redirect that work on the identical floor.

That is the start of agent expertise (AX) within the Copilot app: interfaces the place folks and brokers function collectively. Chat is the place you instruct, focus on, and cause by way of ambiguity. Canvases are the place that intent turns into seen work you possibly can examine, steer, and confirm.

Brokers that may solely counsel code go away you do a variety of the work. To be more practical, brokers must run code, examine outcomes, check modifications, and iterate, with out touching manufacturing.

Cloud and native sandboxes for GitHub Copilot give brokers a bounded place to behave. Select the place Copilot runs—in your native machine or within the cloud—and start unlocking agent-driven workflows whereas prioritizing safety and enterprise coverage enforcement, and with out native useful resource constraints.

With native sandboxing, Copilot runs in an remoted setting straight in your machine, with restricted entry to filesystems, community connectivity, and system capabilities. Native sandbox insurance policies will be centrally configured and enforced.

Within the cloud, every sandbox runs in a completely remoted, ephemeral Linux setting hosted by GitHub. Organizations outline their very own insurance policies. From the cloud, you possibly can decide up Copilot periods anyplace, on any gadget, with distant management.

Code evaluate that scales with agentic output

As brokers produce extra pull requests, the stress on code evaluate compounds. Copilot code evaluate brings an adaptable, agentic system to filter by way of the noise, permitting you to focus your power the place it issues most whereas Copilot conducts code critiques.

Now you can prolong Copilot so each evaluate displays your individual requirements, inside programs, and engineering context through customized agent expertise, MCP server connections, and configurable actions workflows.

Copilot code evaluate now presents medium tier evaluate, which routes pull requests to a higher-reasoning mannequin for higher precision and recall. Admins can set tips for particular person repositories to “low” or “medium.” This allows you to assign lighter, cost-efficient fashions for low-risk code and save extra sturdy mannequin use for repos with increased impression.

The /security-review ability provides Copilot a devoted path for security-focused analysis. The /rubberduck ability is now typically obtainable to make use of a number of mannequin households to critique your implementation and discover novel points.

And in case you’re engaged on Azure DevOps, now you can use Copilot code evaluate natively. Get the identical one-click evaluate, inline feedback, and committable repair strategies you count on, and admins can allow code evaluate on whichever repos they need.

The identical agentic capabilities work throughout the terminal, the cloud, and even your individual instruments, on the identical basis.

Now you can construct your individual instruments with the GitHub Copilot SDK. Now typically obtainable in Node.js/TypeScript, Python, Go, .NET, Rust, and Java, it exposes the identical agentic runtime that powers the Copilot app. In case your staff wants an inside code evaluation device, a customized release-notes generator, or an agent embedded in a help workflow, you construct it on the identical basis as an alternative of wiring collectively a bespoke stack. One runtime, many surfaces.

For builders preferring to work within the terminal, Copilot CLI now has a redesigned interface, voice enter, and scheduled duties to maintain you there.

Copilot CLI has a redesigned TUI in /experimental mode with tabbed entry to drag requests, points, and gists from the terminal. Voice mode makes use of on-device speech-to-text, so audio by no means leaves your machine. /each schedules recurring prompts and background duties.

Cloud automations let brokers run on a schedule, reply to GitHub occasions, open points, and go away feedback. By default, the cloud agent asks permission earlier than every write motion. Swap to autopilot upon getting established belief.

Engineering doesn’t finish with writing code. It contains submitting the difficulty, kicking off the dialogue, and replying to reviewers. Copilot cloud agent can now deal with each a type of steps.

Reminiscence++ and /chronicle give Copilot continuity throughout gadgets and over time. Question context from periods began within the app, CLI, VS Code, or on GitHub.

Accomplice-built agent apps combine with GitHub Copilot to assist automate duties, generate code, analyze context, and execute actions. Use your favourite instruments with out leaving GitHub. Assign points to new brokers that suit your workflow. Companions embrace LaunchDarkly, Shiny, Amplitude, Sonar, Endor Labs, Octopus Deploy, Packfiles, PagerDuty, and Miro. Begin utilizing these agent apps as we speak. And be a part of the waitlist so your organization can even carry its personal agent apps to GitHub.

What we’re constructing towards

Skilled software program calls for judgment, verification, and accountability. That’s the reason the GitHub Copilot app, sandboxes, code evaluate, automation, context, and associate ecosystem are coming collectively as one system: brokers can do extra of the work, whereas builders maintain management of high quality, coverage, and supply.

As agentic workflows develop throughout GitHub, from repository creation to drag request exercise and API utilization, the platform has to develop with them. We’ll proceed to give attention to availability first. We’re dedicated to hardening these programs so agent-native growth is quick, obtainable, and dependable sufficient for groups to rely on day by day.

GitHub is the place that system lives, as a result of it’s already the place the code, the critiques, the problems, and the groups are.

Let’s construct.