GitHub Universe 2026: Copilot Fusion Launches as Agents Start 40% of PRs

GitHub opened Universe 2026 in San Francisco on 28 October with a product that picks models for you. Copilot Fusion routes each step of a coding task to one or more models, and GitHub says its top tier scored within two points of Claude Opus 5.5 Max on the Artificial Analysis Coding Agent Index at half the cost. The same day, the Octoverse 2026 report said agents started more than 40% of pull requests on GitHub in September 2026, up from 17% in March. I run CI for a living, and that second number is the one I would pin to the wall.

What happened at GitHub Universe 2026

The keynote post, written by Jay Parikh, Microsoft EVP of CoreAI, framed the whole event around one line: "Completed work, not more code, is the output that matters." The product list behind that line:

• Copilot Fusion, a multi-model orchestration layer. It looks at a task's needs across reasoning, code generation, debugging and tool use, then decides which model or mix of models runs each step. The Fusion docs describe the tiers.
• Copilot projects, which run agent work across several repositories in managed, parallel sessions, available in the GitHub Copilot app from 28 October.
• Session continuity, so one cloud agent session can be picked up from the app, VS Code, the CLI or the web.
• Copilot automations, repo-level AI workflows that run on events, schedules or by hand, and can now propose changes to themselves for human approval.
• Copilot code review Max, a deeper review tier that GitHub says posted the highest F1 score on ReviewBench for high-severity issues. GitHub did not publish the figure.
• Copilot included in GitHub Enterprise. Enterprise customers get every Copilot feature without a separate subscription and pay for the AI credits they consume. It is sold through GitHub Sales for now, with self-serve "coming soon"; existing Copilot Business and Enterprise customers can stay on their plans.

GitHub also disclosed that it added 4 million CPU cores, 160 PB of fast storage and 50 Tb/s of network capacity in 2026, and that its move to a multi-region Azure deployment should finish early next year. Visual Studio Magazine covered the infrastructure side in detail, and GitHub's engineering blog has a longer post on Git infrastructure for agent-scale development.

The numbers

Octoverse 2026 covers 1 October 2025 to 30 September 2026 and includes private repositories for most metrics.

Why it matters

Two things changed, and they are linked. The first is volume. Actions jobs grew more than threefold in a year, to roughly 1,440 a second, and merged PRs more than doubled. That is what you would expect when a large share of the authors are agents that never get tired of opening pull requests. I wrote about what that does to pipelines in CI/CD when agents open the pull requests; Octoverse now puts a platform-wide number on it.

The second is pricing and routing. Bundling Copilot into GitHub Enterprise and billing by AI credits moves Copilot from a seat line item to a metered one. Fusion then makes the model choice for you, which is convenient, and also means the cost per task depends on routing decisions you do not control. GitHub's own data shows why it built this: 61% of model-choosing users already mix models, and no single model handles more than 25% of requests.

What sceptics say

The headline benchmark claim comes from GitHub and Microsoft. The Next Web noted that the Fusion numbers come from Parikh's post and a LinkedIn post by Satya Nadella, which put it at up to 77% lower cost. I have not seen independent testing of either claim. The ReviewBench claim for code review Max comes without a published score.

The 40% figure also needs care. GitHub calls its agent detection a floor, since it cannot see every agent, and "started by an agent" counts drafts that a human may have heavily rewritten. A higher merge rate for agent PRs (90% within two weeks in September, per Octoverse) tells you they got merged, not that they were reviewed well. More PRs is not the same as more working software, which is Parikh's own point.

My view: the infrastructure disclosure is the most honest part of the event. You do not add 4 million cores in a year for a demo. Whatever you think of the productivity claims, the load is real.

What it means for engineering and DevOps teams

• Budget for metered AI. If you are on GitHub Enterprise, ask your account team how AI credits are priced and capped before anyone enables the bundle. Put a spend alert in place on day one, the same way you would for a new cloud account. The FinOps for AI playbook applies here.
• Capacity-plan CI for agent traffic. If your runners, caches and artifact storage were sized for human PR rates, assume a step change. Rate-limit agent-triggered workflows and keep expensive jobs behind labels or approvals.
• Treat automations like code. Self-improving automations still produce config that needs review. Require the same CODEOWNERS and branch protection for their changes as for any workflow file.
• Ask how Fusion routes. For regulated code, find out which models can be selected, where they run, and whether you can pin or exclude a model per organisation.

FAQ

What is Copilot Fusion?

Copilot Fusion is GitHub's multi-model orchestration layer, launched on 28 October 2026. It evaluates each step of a task and chooses which model or combination of models executes it. GitHub says the Fusion Max tier comes within two points of Claude Opus 5.5 Max on one coding-agent index at half the cost.

Is Copilot now included in GitHub Enterprise?

Yes, as of 28 October 2026 GitHub Enterprise customers can use all Copilot features without a separate Copilot subscription, paying for AI credits used. It is available through GitHub Sales first, with self-serve coming later, and existing Copilot Business and Enterprise plans continue.

How many pull requests do AI agents open on GitHub?

According to Octoverse 2026, agents started more than 40% of pull requests in September 2026, about 110 million that month, up from 17% in March. GitHub says this is a lower bound because some agents are not detectable.

Related: CI/CD in 2026: Agents Opened 1M PRs in 5 Months, Coding Agents in 2026: Cursor at $2B Then Sold to SpaceX and How to Run an AI Coding Agent in GitHub Actions Safely.