Skip to content
Innovation & Growth Talking point

One CNCF maintainer says AI moved his open source work from writing code to reviewing it

Jorge Castro of the CNCF says the maintainer who presses merge still answers for what agents wrote.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
One CNCF maintainer says AI moved his open source work from writing code to reviewing it

Graphic: WebPulse, drawn from the sourced facts in this story.

In brief
  • Jorge Castro of the CNCF said AI agents shifted his open source effort from writing code to reviewing it, and one human still approves.
  • He said expertise still decides what gets merged, even when contributors donate agent time, because someone must understand what was sent.

Jorge Castro, who works on open source community and developer relations at the CNCF, made a simple point on the GitHub show. In his own project, AI agents did not remove the work of open source. They moved it. The effort now sits in reviewing code rather than writing it, and the person who approves a change still carries the blame if it goes wrong. He spoke in a GitHub Universe 2026 session on agents in CI/CD, the automated systems that build and ship software.

What was said

Castro runs Project Bluefin, a Linux operating system built the way teams build applications. He described a setup called hive that coordinates agents. Contributors connect their own agents, which write changes and send them in as pull requests, meaning proposed code changes. A separate review agent comments on each one. By the time Castro looks, he said, five agents and two humans have already reviewed it.

The effect, in his words: "All of a sudden it's all about reviews, not so much about implementation." He said he finds himself drowning in review load. So he now works in batches. He picks one area of the code, looks at every issue and pull request there, and uses another agent to pull them together.

Castro said humans still triage and design the software. Automated triage had not worked well for him. He also said people must understand what they submit, even when they donate agent time. If someone cannot explain what they sent, he will not merge it. On accountability he was blunt: "Even though I do love all this kind of stuff, at the end of the day, it's my neck on the line and I'm the one that pushes the button."

Why it matters

Our reading: a manager who counts AI as saved coding hours may be counting the wrong thing. Castro's account suggests the scarce resource is the approver's attention. More agents produce more changes, but only a person with real knowledge can decide which ones to trust.

That has a training consequence. Judging code needs the experience of someone who could have written it. Castro also said that counting pull requests no longer tells him much. He looks for ways to reward humans who step in and override an agent's decision. Teams that buy or run agent-assisted work could ask the same questions. Who is named as the approver? How much review can that person really do in a day?

The other side

Castro offered no data on whether five agent reviews and two human reviews catch more mistakes. His fix for review overload was more agents, which adds layers of its own. He described his project as an experiment, so this is one maintainer's account, not a measured trend.

He was also candid about the risk. Asked whether AI is good or bad for open source, he said the answer is yes. He told the audience he is terrified of shipping a fully agentic system to live devices. The talk left open how a maintainer team keeps up as the queue grows.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from