How Many CTOs? Is Right About AI-Generated Code, PR Overload, and 98% Uptime
Key Takeaways
- Audit quality gates for AI-speed development, not human-speed habits.
- Run automated checks before merge, especially when code or tests are AI-written.
- Treat AI review settings, attribution, and bypass rules as engineering infrastructure.
Why it matters
- ProductProduct leaders need faster delivery without normalizing reliability regressions that users will feel immediately.
- InvestorsInvestors should watch whether AI-native software teams pair output gains with measurable quality controls.
The new engineering bottleneck is not writing software. It is proving that software still deserves to ship.
I have a small superstition about pull requests. If one arrives with a diff I can read over coffee, I trust my future self to remember why it exists. If five arrive before the coffee cools, I start reviewing the calendar instead of the code. The strange thing about AI coding tools is that they did not just make software faster to write. They made an old social contract visible: review was never only about correctness, it was about pacing.
The speed limit was never written down
How Many CTOs? frames the problem neatly in its Engineering Culture listing for S2E43, where Brad and Scott recap Gergely Orosz’s State of the Tech Industry talk and focus on what broke and what comes next. Their argument is that machine-generated code is now swamping pull requests and code review, revealing that many quality gates were designed around the speed of human authors. That is a cultural diagnosis disguised as a workflow issue.
The same How Many CTOs? listing says the episode also discusses how 98% uptime is becoming a norm teams can no longer comfortably accept. Those two ideas belong together. A team can generate more code than ever, but the user does not experience velocity. The user experiences the checkout page, the login flow, the alert that did or did not fire.
The bottleneck moved from typing to trust
DevPro Journal, citing GitHub’s Octoverse 2026 report, says that by September 2026 more than 40% of pull requests on GitHub began with an AI agent’s first draft. That figure matters because it changes the default assumption of a review. The reviewer is no longer asking only whether a colleague made a reasonable tradeoff. They may be asking whether anyone on the team fully understands the tradeoff at all.
DevPro Journal also cites Veracode testing of 18 AI models, published Oct. 29, 2026, that found a 62% average security pass rate, meaning 38% of the coding tasks produced vulnerable code. The encouraging reading is that AI is useful enough to be everywhere. The sober reading is that usefulness is not the same as fitness for production. DevPro Journal adds that AI-assisted code review grew more than 90% in a year, but says it should surface issues for people rather than approve changes by itself.
Quality after the fact is not a gate
TestMu AI’s State of AI in Quality Engineering 2026 survey describes a familiar compromise: AI writes much of the code and many of the tests, while teams often verify the result later with a person or another model. That is not a quality gate so much as a quality cleanup crew. It may work when the flow is slow. It starts to look fragile when generation becomes cheap and constant.
The TestMu AI survey reports that 81% of teams have merged AI-written code or tests they did not fully understand, and 19% do that regularly. It also says 44% of teams saw more escaped defects or production incidents as AI-generated code grew. Most strikingly, TestMu AI says only 41% of teams run any automated check on AI-written code before merge. If How Many CTOs? is right that 98% uptime is starting to sound inadequate, then shipping first and understanding later is not a harmless habit.
The new gate is programmable, not ceremonial
GitHub’s own Changelog points toward one practical shift. GitHub says customers can now manage Copilot code review settings through REST and GraphQL APIs, from enterprise policy down to individual repositories. It also says teams can identify Copilot reviews and comments through explicit attribution fields. That sounds administrative, but it is actually philosophical: the organization can begin treating AI review behavior as infrastructure, not office etiquette.
This is where engineering leaders have to resist the nostalgic answer. The answer is not to make every engineer read every generated line with heroic patience. The answer is to decide which checks must happen before merge, which risks require a human owner, which AI comments count as advisory, and which repositories deserve stricter policy. If code creation is becoming abundant, judgment has to become more explicit.
The uncomfortable question is not whether AI-generated code is good or bad. It is whether our teams were ever as disciplined as our rituals made us feel. Pull requests gave software organizations a humane speed limit, and AI has found the gap in the fence. For readers building software now, the useful move is to audit the gate, not the hype: what must be understood, tested, attributed, and owned before production gets another change? If 98% uptime is no longer enough, what are you willing to slow down before your users slow you down for you?
Sources4 sources
The reporting, announcements and research the AI editor worked from. Links open the original publisher.
