When AI churns out code, does the bottleneck shift to verification?

How much has code change increased?

As coding agents (AI that writes and fixes code in place of people) spread, the volume of code change itself is changing. The most notable case this week was news that Anthropic had replaced its version control system. By this site's count, the keyword "Scion" was covered by 4 news outlets and 1 community post this week.

According to a GeekNews summary source, when source control, builds and CI (continuous integration, the process that automatically builds and tests changes) became bottlenecks because of AI coding, Anthropic built its own version control system, Scion, and moved its monorepo off Git in five months. The same summary source said daily merges grew from an average of about 2,500 in May 2026 to more than 30,000 in September, and that CI alone performs about 25 million checkouts a day. The collected material includes no official source, so this should be read as a community-relayed account.

If these figures hold, then within a few months the systems that store, build and check code hit their limits before the speed of writing code did. For practitioners, it is a signal that when assessing the effect of adopting agents, they should account not only for development speed but also for the load the repository and CI must handle.

How much is being handed to AI?

On the code-writing side, the scope handed to AI is widening. According to an Ars Technica report on October 13 source, Epic Games is shifting development of its delayed Marvel and Star Wars shooter games to an "AI first" approach, and the report said there is "no more handwritten code." The input material contains no official company announcement.

On October 15, Epic released a preview of its revamped PC game launcher. According to The Verge and Ars Technica sourcesource, the new launcher supports Linux, and Epic said it built the launcher with vibe coding (a way of having AI write code from natural-language instructions). Epic claimed the new launcher is much faster, and Ars Technica reported that Epic said it is faster than Steam source. The comparison conditions are not in the material. Epic noted that the preview app may have bugs sourcesource. The news was also shared in the community source, and by this site's count it was covered by 14 news outlets and 1 community post this week.

Tooling is moving in the same direction. According to a GeekNews write-up source, the open-source coding agent OpenCode 2.0 introduced "Code Mode," which turns MCP servers and OpenAPI endpoints into script functions, as well as background agents. GitHub said it revamped HydraFusion, which coordinates multiple models to produce a single response, to use medium reasoning effort for drafts and high reasoning effort for review, and added parallel critique models to the Critique workflow source. The feature is still a research preview source.

A design that drafts lightly and spends more compute on review can be read as a sign that tool makers, too, are starting to put more weight on "checking" than on "writing."

Who verifies all the extra code?

As change volume grows, it becomes hard for people to read every change carefully. The two features GitHub made generally available on October 15 target this point.

According to GitHub's announcement source, "AI Scan for pull requests" finds vulnerabilities in pull request changes in areas CodeQL does not yet cover, such as PHP, Shell/Bash, Terraform configuration (HCL) and Dockerfiles. Results are advisory, so they do not fail code scanning checks or block merges, and the scan does not run on PRs from forks or created by Dependabot source. "Agentic autofix" explores relevant files to create a fix, reruns CodeQL to confirm the alert is resolved, and then opens a draft PR for review source. Up to 25 alerts can be selected to be fixed together in one PR, but only alerts on the default branch are eligible source. Costs come out of AI credits in the Code Security budget and GitHub Actions minutes source.

Together, the two features form a flow in which "AI finds, AI fixes, tools recheck, and a human approves." This can be seen as a structure in which the human role shifts from writing code directly to guarding the final gate. It should also be noted that verification costs now show up as new line items: CI minutes and AI credits.

Has the bottleneck really shifted to verification?

The signals are clear, but some points still call for caution. The only concrete figures showing a surge in change volume come from the Anthropic case, and even those are from a community summary source, not an official source. Epic's "AI first" shift has also been conveyed only through press reports source.

Verification tools do not replace human judgment either. AI Scan results are advisory information that does not block merges source, and Agentic autofix ultimately produces a draft PR awaiting human review source. On the open-source side, how to handle AI-written contributions appears to be emerging as a topic of discussion. By this site's count, the keyword "Linus Torvalds Claude Code" was covered by 4 news outlets and 1 community post this week, and "QGIS AI contribution ban" by 1 news outlet and 2 community posts. Details are not in the collected material.

In short, on the question of whether the bottleneck is shifting from writing to verification and integration, signals are coming from both infrastructure (Scion) and tooling (GitHub). How far it has shifted, however, is something each team must check with its own metrics. Looking at merge counts, CI wait times and the number of PRs awaiting review together will reveal where your team's bottleneck lies.