GitHub Copilot’s Agent Era: What Organizations Need to Know

By Or Shohat, Senior Cloud and AI Architect at UBTECH

At Universe 2026, GitHub made a series of announcements about Copilot: cloud sessions that move between interfaces, automations, a memory and context layer, a way to choose models, security scanning, and a change to the licensing model. Each one is a feature in its own right. Together, they point in a clear direction: Copilot is changing from an assistant that responds to prompts in the IDE into a system of agents that run in the background, over longer periods, and across several tools.

I help development teams work with Copilot, and the same four questions come up in every organization: Where does it run? What does it know about our code? How much does it cost? Who controls it? These questions are a good way to read the announcements.

One Session, Multiple Interfaces

The Connected Cloud Experience lets you start a task on github.com, in GitHub Mobile, Copilot CLI, the Copilot app, VS Code, Slack, or Teams, and continue it in another interface within the same session. By default, every cloud session runs in a sandbox on GitHub’s infrastructure. When a session shuts down, a snapshot is saved so you can resume from it. In other words, the agent no longer depends on the developer’s machine.

In VS Code, a component called Agent Host separates sessions from the editor window. Sessions keep running in the background even after the project is closed, including on a remote machine over SSH. There is also a new option (in Preview) to run several models in parallel on the same task and compare the results.

The Copilot app adds two capabilities in Public Preview:

  • Projects takes a goal and context (repositories, mockups, Slack threads, meeting summaries) and launches several agents as needed.
  • Computer Use lets Copilot work with desktop applications on macOS and Windows once the user approves.

Automations

With Copilot Automations, you define a task, a trigger, and the tools Copilot is allowed to use. Copilot then runs the task every time the trigger fires. GitHub announced four changes at Universe.

The first change adds more triggers and actions:

  • events from external systems such as Jira, Sentry, and PagerDuty through the REST API
  • 24 new triggers within GitHub, including code scanning, secret scanning, and Dependabot alerts
  • 20 new tools
  • automation chaining
  • support for MCP servers

You can also set AI Credit caps per run, day, week, or month.

The second change is shared automations. These belong to the repository, not to an individual user. They run under Copilot’s identity, are billed to the organization, and keep running after their creator leaves. The third change lets you manage cloud automations from within VS Code. The fourth is an automation that reviews another automation. Once a week, it looks for signals such as labels that were corrected by hand or code review comments that were not accepted. It then suggests changes to the prompt, the tools, and the triggers. A change takes effect only after it is approved.

From PR to Merge

Agent Merge monitors a pull request and fixes whatever is blocking it, mainly failed workflows and merge conflicts. It is now also available in the PR’s merge box through the “Watch and fix” button. When combined with auto-merge, the PR is merged once all configured requirements are met. In organizations where PRs wait for days between red CI runs and conflicts, this can cut the time spent following up. Organizations that turn it on by default should make sure their branch protection and review rules are set up correctly.

GitHub Intelligence: Context and Memory

GitHub Intelligence is a context layer built on the information that is already in GitHub. It has three components.

Copilot Memory is now generally available (GA) on paid plans. It keeps personal preferences and repository conventions across sessions and interfaces. GitHub’s data from the Preview period shows that in CLI sessions that used memory, 93% followed developer preferences more closely and 42% worked more efficiently.

Dreaming (in Public Preview) goes back over completed work. It updates Memory, suggests changes to instructions and skills files, and turns comments that keep coming up in code review into standing guidelines. A few practical points:

  • Changes to repository files are submitted as pull requests.
  • The feature must be enabled in settings.
  • Each run uses AI Credits.

The third component is cross-repository Code Search. It includes Semantic Search and a /research command for questions that span several services.

Here, as with automations, the pull request is the control point: any change to guidelines the team maintains waits for human approval.

Model Selection: Copilot Fusion

Fusion is now GA. Instead of choosing a model, you choose a reasoning level: Standard, High, or Max. For each prompt, Fusion decides which models to use. It picks from Anthropic and OpenAI models such as Claude Opus 5.5, Claude Sonnet 5.5, and the GPT-6 family. Sometimes it uses a single model. Sometimes a lower-cost model writes a draft that then goes through a quality check. Sometimes one model writes and another reviews.

On the Coding Agent Index from Artificial Analysis, Fusion at the Max level scored 60, two points below Claude Code with Opus 5.5. Its average cost was $5.59 per task compared with $11.42, and it used about half as many tokens. GitHub also reports that tasks finish about 40% faster. Each model used is billed at its standard rate, and paid plans currently get a temporary 30% discount. Measure the cost on your own team’s tasks before you draw conclusions from the benchmark.

Security and Control

Code scanning for AI applications (in Public Preview) finds vulnerabilities that are typical of agents:

  • indirect prompt injection
  • exposure of sensitive data without access controls
  • more permissions or autonomy than the task needs

Prompt injection ranks first on the OWASP Top 10 for LLM Applications. The scanner runs on every pull request alongside CodeQL. It flags code paths where untrusted content reaches a model that can access tools or data, and it suggests fixes. It requires a GitHub Advanced Security license. For now it is available only on github.com, not on GitHub Enterprise Server.

Most of the announcements also come with controls:

  • By default, a cloud session is limited to a single organization, with sign-in through SSO.
  • In a repository where Copilot has no write permission, it can only read.
  • Using certain tools and controlling applications requires approval.
  • Fusion uses only models the organization has approved.
  • Admins can require or block plugins in VS Code.

The controls exist, but setting them up is the organization’s responsibility.

Licensing and Pricing

GitHub Enterprise now includes Copilot access for every licensed user. For now, this option is available on GitHub Enterprise Cloud to organizations that work with a GitHub or Microsoft account team. The seat is included, but AI usage is metered and billed to the organization’s budget from the first request. There is no monthly allowance per user. In practice, managing licenses is replaced by managing usage and budgets. Users without an Enterprise license can get access through Copilot Standalone for $10 per user per month. GitHub has also renamed the GitHub Copilot SDK to the Microsoft Copilot SDK. The harness built on it is now the default for new sessions in every Copilot interface.

The Bottom Line

The Universe 2026 announcements fit together: cloud sessions, automations, memory, model selection, handling PRs all the way to merge, and security controls. Many of them are still in Preview, and the move to usage-based billing means costs need to be monitored. To adopt them, an organization will need to set policies, budgets, and review processes, and measure the real benefit. A team that wants to try this direction can start with a small pilot, such as one automation or Agent Merge in a single repository, and expand based on the results.

Originally published on Geektime