GitHub Universe 2026: The goal isn’t more code anymore

We recently got back from GitHub Universe in San Francisco, and we’re still buzzing. Our team from Platform, Quality, and Security Engineering spent the week in sessions, at demos, and in hallway conversations with other people building on GitHub. Hosting our happy hour at Shotski’s Alpine Lodge was a highlight. It was a relaxed evening of good company, and the kind of conversations you only have in person: what’s working, what’s messy, and what everyone wishes they’d known sooner. We also caught up with our partner teams at GitHub, LaunchDarkly, and Atlassian, and we’re grateful to everyone who stopped by. We’ll do it again next year, so save a spot on your calendar.

A theme ran through all of it. The conversation had moved on from “look how fast AI writes code” to a harder question: how do you turn all that code into finished work? GitHub put it plainly. The output that matters isn’t more code. It’s completed work: a shipped feature, a bug fix, a security fix.

Below are the five announcements we think matter most, what we heard from the people we talked with, and where we’d start. If you couldn’t make it to San Francisco, consider this your seat at the table.

The 5 things worth your attention

1. The Copilot app turns Copilot from a helper into a teammate

It takes work from issue to merge in one place. The headline feature is Projects, now in public preview: give Copilot a goal and it runs agent work across multiple repositories in parallel, with one place to track progress and steer. Automations run in the cloud on a schedule, an event, or a click, and they can now learn from their own results and propose improvements to themselves. A triage automation, for example, can learn from whether people change the labels it applied. You review each proposed improvement before it takes effect.

2. Fusion stops the model-picking game

Developers shouldn’t have to guess which model fits which task. Fusion, now generally available, decides for them. GitHub reports that Fusion Max scored within two points of Claude Code running Opus 5.5 Max on the Artificial Analysis Coding Agent Index, at about half the average API cost per task and about 40% less time. Those are GitHub’s reported results, so run Fusion on your own work before you bank the savings.

3. Code review Max is built for the flood

More agent-written code means more pull requests, and someone must review them. Max is a deeper review tier, now in public preview, sitting beside Lite and Balanced, and GitHub says it posted the top F1 score on ReviewBench for high-severity issues. It also takes longer and uses more AI credits per review. Use the lighter tiers for everyday changes and save Max for the ones that scare you.

4. Copilot now comes with GitHub Enterprise

Enterprises working with a GitHub or Microsoft account team can now get Copilot included with GitHub Enterprise for every licensed user, with no separate Copilot subscription to buy. Copilot moved to usage-based billing in June 2026, and that doesn't change. What changes is how you manage it: Copilot access comes with the license, and you pay for the AI credits you use against your enterprise budget. It's available through GitHub Sales today for GitHub Enterprise Cloud, with self-serve availability coming in stages. Existing Copilot Business and Copilot Enterprise customers can keep their current plans. This is the quiet change that matters most. The question shifts from "how many seats?" to "who is spending what, and on what?"

5. Git itself is being rebuilt for agents

GitHub is re-architecting its Git infrastructure for agent-scale traffic while the platform keeps running. The work is in progress, and GitHub reports up to 35x higher write throughput in internal benchmarks. The reason: pushes on GitHub grew almost 5x in a year. You have nothing to adopt, and GitHub says the new design is built to preserve branch protections, required reviews, and audit logs. But it tells you how much load GitHub expects, and it’s a good prompt to ask whether your own CI pipelines could absorb the same growth.

What we heard on the ground at GitHub Universe

The announcements were half the story. The other half came from conversations with engineering, platform, and security leaders throughout the week. A few themes came up again and again.

Copilot is on, but the structure isn’t

Many teams have rolled out Copilot and are still working out how it should be used: which workflows it fits, how to bring teams along, and how to tell whether it’s paying off.

Cost is becoming a real conversation

As usage grows, people want to know where their credits go and how to get more from them. Fusion and the new Enterprise model make that question more pressing, not less.

Security has to keep pace with agents

Security leaders asked how to put guardrails around AI agents, and how to cut alert noise so real findings don’t get lost.

Modernization is still in flight

Several teams are mid-migration, consolidating source control and modernizing pipelines in regulated environments, while trying to add AI on top.

Great announcements. Now what?

Here’s the thing we kept coming back to all week: none of this is hard to turn on. The hard part is what happens next.

Picture the good version. An agent picks up a dependency upgrade across a dozen repos while your team is in standup. It opens the pull requests. Copilot flags the high-severity issues. Your tests and required checks decide what merges. An automation triages the follow-up issues, and your people set direction and sign off on what ships. That’s a great Tuesday.

Now picture the other version. Credits burn faster than anyone can see. Agent-written code slips through with less scrutiny than human-written code would get. Your pipelines buckle under the extra volume. And if you work in a regulated industry, someone asks, “Can you show what ran, who approved it, and what it cost?” and the room goes quiet.

The difference between the two is the groundwork you lay now. Here’s where we’d start:

Give the money an owner

Name who owns AI credit budgets and reporting before you widen access or move to the new Enterprise model. Budgets cut both ways: set them too loose and spend surprises you; set them too tight and work stalls. On the new model, Copilot Chat, code review, and Copilot in the cloud pause when the budget runs out, while code completion keeps working. It’s a small decision that prevents a big surprise.

Start small and real

Pick one low-risk change that spans several repos, like a dependency upgrade that needs code changes, not just a version bump, and run it as a pilot in Copilot Projects. Projects is still in public preview, so a contained pilot is the right way to try it. You’ll learn more from one real run than from ten demos.

Let the reviewers compete

Turn on Copilot code review for one team and compare Lite, Balanced, and Max on your own pull requests. Track what each tier catches against what it costs: Max takes longer and uses more AI credits, and every Copilot review also uses GitHub Actions minutes. Find out where the deeper tier earns its keep.

Hold agents to the same bar as people

Make sure your required checks, branch protections, and security scanning treat agent-written code exactly like human-written code. No fast lanes. That includes Copilot’s own approvals, which count toward your required reviews once you turn them on. Decide which file paths, if any, Copilot may approve before anyone enables it.

Look back in 30 days

See what worked, what surprised you, and what it cost, then write the policy for wider rollout, including which newer capabilities to enable, limit, or leave off for now.

Wherever you are on your GitHub journey, we’re happy to compare notes. As a GitHub partner, we help enterprises move to GitHub and modernize CI/CD with GitHub Actions, build security in with GitHub Advanced Security, and get real value from Copilot. See what we offer through our GitHub partnership, or just tell us what you’re working on. We’d love to hear about it.

This article was developed with contributions from Joel Gerbino, Director, Quality Engineering.