Too Much Code: AI Output Is Now Breaking Software's Plumbing

The Moment Everything Changed

For two years the question about AI coding was whether the machines could write good software. This week the question changed: what happens when they write more software than the rest of the system can absorb? Anthropic said its own repository had outgrown git, PostgreSQL pushed back a major release because AI reviewers found more bugs than people could fix, and Linux kernel maintainers dropped the rule requiring contributors to admit they used AI. Taken separately, each is a niche engineering story. Taken together, they show that the bottleneck in software has moved from writing code to everything that happens after the code is written.

Background

The infrastructure modern software runs on, from version control to CI pipelines to code review queues, was designed around how fast humans work. Git, created by Linus Torvalds in 2005, assumes developers commit a few times a day and that a repository can be cloned in full to a laptop. Review processes assume that a reviewer’s attention is the scarce resource and that the volume of changes is roughly limited by how many engineers a project employs. Those assumptions held for twenty years because human typing speed acted as a natural limit on throughput.

Agentic coding tools removed that limit over the past year. GitHub COO Kyle Daigle said in April that commit volume on the platform would likely reach about 14 billion in 2026, roughly fourteen times the 2025 total, according to DevX. GitHub logged ten service-degrading incidents in April alone, and a Microsoft spokesperson said the “spike in agentic development” had tested the platform’s limits. Developers noticed too, and HashiCorp co-founder Mitchell Hashimoto complained that GitHub “is no longer a place for serious work” when it locks users out for hours.

What Happened

The clearest data point came from Anthropic, which published an engineering post on October 13 describing why it replaced git for its main monorepo. The company said it averaged about 2,500 merged code changes per day in May 2026 and more than 30,000 per day by September, a twelvefold jump in four months, as reported by The Register. It runs on the order of 25 million code checkouts per day for CI, which it said would amount to more than 600 PB of git traffic daily if fetched naively. Its main branch history is now close to 100 GB. Anthropic added that it doesn’t expect growth to slow, because “the latent demand for merging code is practically infinite.”

Its answer is Scion, a centralized version control system that a seven-person team built, largely with Claude, and migrated the monorepo onto in five months. The server handles every commit, rebase and merge, and clients keep only a bounded local cache, which means Scion has no offline mode. Anthropic shrugged that off because Claude, its heaviest user, is always online. The company also built a code review interface called Antique, a code search tool and a source-control permissions layer. It says Scion won’t be open sourced and isn’t meant as a general git replacement.

Three days later PostgreSQL, one of the most conservatively run projects in open source, delayed version 19 from early October to around November 19. Developers pulled eight to ten significant features, including SQL/PGQ property-graph queries, and deferred them to PostgreSQL 20, according to The Register. The cause was LLM-assisted code review that kept turning up edge-case defects in features that had already been committed. Committer Tom Lane warned that shipping SQL/PGQ in v19 would mean post-release bugs that could not be fixed until v20.

The Linux kernel went the other way on disclosure. At its Maintainers Summit the community agreed to deprecate the “Assisted-by” tag, which had entered the kernel documentation only in January, The Register reported. Red Hat’s Jeff Layton argued that compliance was patchy and that the tag served as “free advertising for proprietary LLM companies.” Greg Kroah-Hartman had already said in August that LLMs are now good enough to find kernel bugs.

Why It Matters

The common thread is that AI has made producing code, and producing criticism of code, close to free, while the systems for merging, reviewing and releasing it still run at human speed. When generation is effectively unlimited, every downstream step becomes the constraint: source control, CI compute, reviewer attention, release schedules. Anthropic’s engineers said their main job had become telling Claude “no” when agents broke core system constraints to satisfy a pattern. That describes a change in what software engineering is, with the human acting as gatekeeper rather than author.

The PostgreSQL delay shows the same pressure from the quality side. AI reviewers don’t get tired, so they surface every minor defect at once, and each one carries what committer Peter Eisentraut called a fixed overhead for humans to handle. More scrutiny should eventually give more robust software, but in the short term it means slower releases and painful choices about what to cut. Microsoft’s record 1,411-vulnerability Patch Tuesday this week, which the company attributed to AI-assisted bug discovery, according to BleepingComputer, is the same dynamic at industrial scale.

There is also a trust gap. Stack Overflow’s latest survey of 30,903 developers found that about 80 percent of daily AI users spend at least an hour a day with these tools, yet only about 7 percent trust AI output for important decisions, as summarized by The Register. Developers are writing far more code with tools they don’t fully trust, while the institutions meant to check that code are being overwhelmed. That combination is why provenance, review tooling and infrastructure capacity are now strategic problems rather than housekeeping.

Expert Perspectives

Not everyone reads Scion as revolutionary. Developer and commentator Simon Willison said it “sounds to me like a reinvention of centralized source control along the lines of Subversion,” and guessed that the arrival of Fable- and Mythos-class models “is a key factor in the increased coding velocity,” writing on his weblog. His point is a useful corrective. Google, Meta and Microsoft built similar centralized systems years ago, and what is new is that one AI lab reached hyperscaler-level code volume with a fraction of the headcount.

On the review side, Eisentraut was candid about the trade-off: “With LLM-assisted code review, these kinds of things can get found much faster,” he said, before adding that “there is a fixed overhead to dealing with each of these, and then you run out of time,” as quoted by The Register. He rejected the idea that AI-written code caused the defects. In his view “the effective quality bar has been raised” after the features were committed, and that is a different and more hopeful diagnosis.

The kernel debate showed the disagreement over accountability most sharply. Developer Laurent Pinchart objected to dropping the tag without a replacement, arguing there is “even more value now” in stating what the project expects from contributors. Linus Torvalds dismissed tool-specific rules outright, saying what matters is whether contributors understand their code, not what tools they use. Both positions accept that AI is now part of kernel development. The open question is whether provenance still needs recording once nearly everyone uses these tools.

What to Watch

The first signal is GitHub’s capacity. Microsoft is reportedly drawing on AWS for emergency headroom while it speeds up its Azure migration, and another quarter of agent-driven growth will test whether the world’s default code host can keep up. If it can’t, expect more large engineering organizations to follow Anthropic onto custom or centralized systems, and expect startups selling agent-scale source control and CI to multiply. The economics of CI compute, now measured in petabytes and millions of daily checkouts, deserve as much attention as model token prices.

The second is open-source release cadence. PostgreSQL 19’s November ship date will show whether AI-assisted review becomes a quality advantage or a recurring source of delay, and other conservative projects will be watching. The third is provenance. Linux has stepped back from mandatory disclosure just as Deezer, YouTube Music and Stack Overflow’s respondents push for more attribution elsewhere, and how code hosts, licensing bodies and regulators handle AI-written code will shape software liability for years.