GitHub Universe 2026 Recap: Fix the process first, then add the agents

Here's the short version of GitHub Universe 2026: the teams getting real results with agents fixed their process first and picked tools second. I spent October 27-29 in San Francisco at Universe for World Wide Technology (WWT), and almost every session I sat through landed on that point from a different direction. (There is a plush squirrel in this story. Stay with me.)

Agents run the loop. People own the outcome.

GitHub opened with scale, and some of the numbers are big enough to be silly. In its Octoverse report, published during the keynote, GitHub counted more than 225 million developers, 34.3 billion commits, and 1.3 billion merged pull requests in the 12 months ending September 30, 2026. Pull requests that begin with an agent's first draft passed 40 percent in September, up from 17 percent in March. That is a big shift in six months.

Mario Rodriguez, GitHub's chief product officer, said an agent's value shows up in the software work it finishes, and he pushed back on the idea that developers get replaced by software factories: Humans set direction. Agents close the loop. I think this is an important split, because it keeps a person accountable for what ships.

GitHub backed the framing with product. The keynote introduced GitHub Copilot Fusion, which coordinates several models to hold quality while cutting cost, and Copilot Automations, cloud workflows that respond to events and schedules. GitHub also described a new Git architecture that separates storage from compute so repositories can absorb agent-scale traffic. Evan Boyle then showed Lever, an open source system that points agents at goals like getting CI to green. People set the goal and the autonomy level, and they make the merge call. His advice: hand the boring work to agents and take the time back.

Fix the process first

The common mistake is starting with the tool. The strongest customer stories started with the process.

Levi Strauss & Co. ran a three-day workshop to map its software lifecycle: four stages, 12 handoffs, six roles. The team then sorted every step into what AI does, what a human does, and what a human decides. Seventeen specialist agents now cover work from requirements to deployment, with approval gates built in. They reported that development time for their "BackPocket" stylist app dropped from nine months to two and a half, with early results showing a 51 percent throughput gain, and AI-assisted pull requests climbed from 26 to 69 percent. They also shared that the tooling was the easy part. Change management took the real work: pre-reads, hands-on training on real code, and office hours.

Trimble found that 436 of 439 open source repositories were out of compliance with policy. They could have spent months building remediation tools or chasing maintainers, but instead spent four days building an automated governance system with GitHub Copilot, and wrote none of the code themselves. The work went into the policy, the edge cases, and making every fix reversible. The system opened issues with exact fixes, deadlines, and a reusable Copilot prompt, so the compliant path became the easy one. All 33 public repositories were compliant within four days.

Governance decides whether any of this scales

Once agents work across many repositories, control becomes the main design problem. A few patterns kept showing up.

• Build a control plane. Marks and Spencer described a curated catalog of approved agentic workflows that engineers can deploy to a repository in one click, with shared configuration and telemetry. The proof of concept raised 123 issues and fixed 77 of them the same day.
• Keep an audit trail. Provation works in healthcare, where every automated decision needs one. They instrumented agent workflows with OpenTelemetry and cut automated AI spend by 90 percent.
• Watch what the code does at runtime. Isovalent ran a live demo where a GitHub Copilot-written change added a dependency that quietly called a payment service. Tetragon, an eBPF-based security tool, flagged the call, blocked it, and then killed the process. Their advice was to move those policies into CI, and treat AI-generated code with some skepticism.
• Write the rules down for agents too. Jorge Castro of the CNCF said agent-generated pull requests overwhelmed the maintainers of Project Bluefin. Instead of banning agents, the team wrote separate contributor guides for humans and agents, added review agents, and kept human approval in place. GitHub's closing keynote added maintainer controls to the mix, including limits on work in progress and the option to disable pull requests.

A panel with leaders from Anthropic, OpenAI, and Google Cloud pointed the same way. They said the industry over-invests in product-specific agent skills and under-invests in verification, observability, agent identity, and security. That is the unglamorous engineering work in between, and it is where the value sits.

What I'm taking back

Of all the great insights to choose from, I think the biggest thing that stood out: Pick one process. Decide where agents act and where people decide. Instrument all of it. 

That is the focus of our AI-native engineering work at WWT, and Universe backed it up.

I've watched this pattern before with DevOps transformations and CI/CD implementations: The tool shows up first, and the teams that get value are the ones that fix how they work before implementing. Agents are the latest round.

Now, as I mentioned earlier, the squirrel. In the closing keynote, GitHub's Martin Woodward had Copilot help him train a decision tree on badge sensor data, shipped a working pedometer on stage, and then trained a badge inside a plush squirrel to tell a throw from a drop. A room full of developers got a machine learning lesson from a stuffed animal. I'd sit through that again.

Before you start your own agent project, ask yourself:

• Which process would you redesign first?
• Where do agents act, and where do people decide?
• What would you see if an agent's code did something unexpected in production?
• Who is accountable when an agent's pull request merges?

Let us at WWT help you answer these questions and plan for the next steps in your AI-native Engineering journey. Connect with us in a complimentary one-hour AI-Native Engineering Briefing - WWT