GitHub Universe 2026 Copilot: Agent Host and Harness
GitHub opened Universe 2026 in San Francisco on October 28, and most of its Copilot news is about where an agent session lives. The Copilot harness is now the default for new sessions in VS Code. A new Agent Host keeps sessions running after you close the window. You can pick up the same session in another Copilot client, including your phone. By the end you will know which of the October 28 Copilot announcements are still in preview and what Copilot remembers once a session ends. Universe runs through October 29, so later announcements are not included.
What did GitHub announce for Copilot at Universe 2026, in one sentence?
GitHub made one harness the default for new Copilot sessions in VS Code and moved sessions into an Agent Host that keeps them running across Copilot clients.
The main source is the VS Code Universe releases changelog entry, one of several posts GitHub published that day. It says the harness, built on the Microsoft Copilot SDK, now runs new sessions in both the editor and the Agents window. GitHub says the switch makes VS Code behave more like Copilot CLI and the GitHub Copilot app.
Microsoft published its own harness announcement the same day. It describes a harness as "everything around the model," from instructions and context to the environment the agent runs in. GitHub's engineers built it, and Microsoft has adopted it company-wide, including in Copilot Studio. Microsoft says it supports more than 20 frontier models plus bring-your-own-key.
How the Agent Host differs from a classic VS Code chat session
A classic Copilot chat lives inside the editor window that started it. The Agent Host docs describe the new model as "a dedicated process that runs agent harnesses and manages their sessions." Several clients can watch and steer one session. The host is the source of truth for its state.
| Local harness (extension host) | Agent Host session | Cloud session | |
|---|---|---|---|
| Where the work runs | Inside one VS Code window | A separate process on your machine, a Dev Container or a remote machine | GitHub's infrastructure |
| After you close the window | The session ends | Keeps running while the host runs | Keeps running |
| Open it from another client | No | Yes, clients reconnect and catch up | Yes: VS Code, Copilot app, CLI, GitHub Mobile, github.com |
| What must stay on | Your editor window | The host process and its machine | Nothing local |
| Copilot connectors | Not supported | Supported with the Copilot harness | Not covered in the connectors docs |
Sources: the Agent Host docs and the VS Code changelog. The connectors row comes from the connectors docs.
The docs add one catch. Tools that a client supplies, such as browser or extension tools, stop working when that client disconnects, even though the session keeps running.
How does Copilot session continuity work?
Cloud sessions start from the Agents window in VS Code and run on GitHub's infrastructure, so they keep going when you close your laptop. Per the changelog, you can continue that session in any of the five clients in the table above. Sessions started from a Slack or Microsoft Teams DM can be continued there too.
The local path keeps the agent on hardware you control. The Agent Host can run on a separate machine over Remote SSH or dev tunnels, where the agent uses that machine's code and dependencies. It keeps working after you disconnect. To steer a machine from your phone, run code agent host --github-environment in the project folder and then open GitHub Mobile. Both setups need the machine to stay on with its Agent Host running.
You can also switch between the editor and the Agents window without restarting a task. Both views show the same live session. The protocol behind this, the Agent Host Protocol, is open. It ships client libraries in five languages, so a third-party client can sync with a running host.
What is Copilot Fusion, and how does model choice work now?
Copilot Fusion is generally available on all Copilot plans. It works in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. You pick Fusion in the model picker and choose an effort level: Standard, High or Max. Copilot then decides which models handle each task. Sometimes one model does the work. In the Cascade pattern an efficient model drafts and a quality check escalates to a stronger one. In the Critique pattern a second model reviews the first model's draft.
GitHub reports that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index. That is within two points of Claude Code with Opus 5.5 Max. GitHub says Fusion Max got there at about half the average API cost per task and in about 40% less time. Billing is each model's standard rate, with a 30% discount on Fusion model costs for paid plans for a limited time.
VS Code also added a preview that runs several models or harnesses in parallel in the Agents window. A Judge model can recommend one result. A Synthesize option merges ideas from the other attempts.
What else can Copilot reach now?
Computer use lets Copilot click controls and edit text in desktop apps on macOS and Windows. It is a public preview in Copilot CLI and the Copilot app. In VS Code it is experimental: turn on chat.agentHost.computerUse.enabled, restart, then type /computer on in a local session. Copilot asks for your approval per app, and organizations can disable the feature through policy.
Copilot connectors are GitHub-managed MCP servers. The launch examples are Slack, Notion and Microsoft Learn. You connect once and reuse the connection in supported Copilot clients signed in with the same account. Connectors are in preview on all plans. On Business and Enterprise, an admin must first allow editor preview features.
Two announcements push work beyond a single session. The Copilot app's Projects (public preview) take a goal plus context and coordinate as many agents as the work needs. Shared Copilot Automations are rolling out on paid plans from Pro upward. Each automation belongs to a repository and runs in the cloud on triggers such as a new issue. It keeps running if its creator leaves.
What does Copilot remember after a session ends?
GitHub's intelligence announcement says Copilot Memory is generally available on paid Copilot plans. It keeps personal preferences and repository conventions across sessions and Copilot clients. Copilot writes and curates these memories itself. GitHub says that in public-preview CLI sessions that used memory, 93% showed better adherence to developer preferences. An opt-in preview called Dreaming reviews past work and can propose changes to skills and instruction files as pull requests. Each dream run uses AI credits.
GitHub's Copilot Memory docs list the limits. Memories expire after 90 days without reinforcement by default. You can view and delete memories, but on GitHub.com you cannot create or edit a single one. In VS Code, memory is a local setting that is off by default. Apps built on the Copilot SDK can turn memory on for a session, but the SDK has no APIs for managing individual memories.
What should you read in the fine print?
Almost everything here comes from GitHub and Microsoft, and much of it is in preview. Connectors and Projects are previews. So are Dreaming and the parallel model comparison. Computer use is a preview in the CLI and the app and experimental in VS Code. GitHub's own computer use docs warn that it can pick the wrong control or type in the wrong place. They advise against "Always allow" for sensitive apps.
The Fusion benchmark comparison and the 93% memory figure are GitHub's own reports, and we found no independent test of either. A local or remote Agent Host session lives only as long as its host process and machine. The docs also note that a saved cloud conversation does not guarantee its cloud environment still exists.
The bottom line
If your team already pays for Copilot, try the Agent Host on a remote dev box for long tasks you used to babysit. Use cloud sessions for work you want to check from your phone. Treat computer use as a preview and keep it away from finance and admin apps. Turn on Copilot Memory in VS Code once you know which repositories should build it up.
Where does agent memory live if you also use Claude Code or Codex?
Say Copilot learns that your repository's integration tests need a seeded database, and Copilot Memory keeps it as a repository convention. On Tuesday you open the same codebase in Claude Code. Claude Code never sees Copilot's memory, so it has to learn the setup again. memU will not carry Copilot's memory across either. What it does is give Claude Code and Codex one shared memory, so outside Copilot the setup only has to be learned once.
| Copilot Memory | memU | |
|---|---|---|
| Which tools read it | Copilot clients, across sessions | One memory shared by supported hosts such as Claude Code and Codex |
| Can you fix an entry | View and delete only, with no per-entry editing on GitHub.com | Entries are readable Markdown files you can open and fix |
| Skills | Dreaming (opt-in preview, uses AI credits) proposes skill and instruction changes as pull requests | Reusable skills are distilled from agent history as Markdown |
| Copilot support | Built in on paid plans | Copilot is not a supported memU host |
If your team works only in Copilot on a paid plan, Copilot Memory is likely enough. It comes with the plan, and in VS Code you turn it on first. memU is for the developer who also runs Claude Code or Codex on the same code and wants those agents to share one memory.
To try the hosted version, which is free, get an API key from memu.so and send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is
YOUR_MEMU_API_KEY.
The self-hosted setup and the list of supported hosts are in the memU GitHub repository. For background, see our earlier post on VS Code as a multi-agent hub.
FAQ
What is the Copilot harness?
It is the layer around the model that supplies an agent's instructions and context. GitHub built it on the Microsoft Copilot SDK. It is now the default for new Copilot sessions in VS Code.
What is the Agent Host in VS Code?
The Agent Host is a separate process that runs agent harnesses and owns their sessions. Sessions keep running after you close a VS Code window. The host can run on your own machine or on a remote one.
Can I continue a Copilot session on my phone?
Yes. Cloud sessions can be continued in GitHub Mobile. You can also steer an Agent Host on your own machine from GitHub Mobile after running code agent host --github-environment there. The machine has to stay on.
Is Copilot computer use available in VS Code?
It is experimental in VS Code on macOS and Windows and needs a setting plus a restart. It is a public preview in Copilot CLI and the GitHub Copilot app. Organizations can disable it by policy.
Does Copilot remember things between sessions?
Yes, on paid plans. Copilot Memory stores repository facts and personal preferences across sessions and Copilot clients. Unreinforced memories expire after 90 days by default.