Did AI Remove the Coding Bottleneck? Where the Work Went (Code Review)
By Matt Cullerton, founder of Mavric and Apso
A TikTok from @agenticengineering asks why engineers work 13-hour days if "AI removed the coding bottleneck." The 13-hour figure comes from a single anonymous post, and the bottleneck did not disappear. Telemetry from Faros AI on 22,000 developers shows the work moved into code review, where median review time rose 441.5% on teams with high AI adoption.
Where does the 13-hour workday claim come from?
From one post on X. Simon Willison quoted it on 20 September 2026, attributing it to voxium (@v0xium), two days before the video. The post says "The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code," and "People are working 12 to 13 hours a day just to press enter." The author had been in a new role at an unnamed big company for half a month.
No headcount, time tracking or named employer backs it up. The video's caption rounds the range to "13-hour days," while the narration keeps "12 to 13 hours a day." The speaker adds his own caveat: "this is one engineer describing one company, so I would not generalize too much from it."
How many hours do software developers actually work?
Sourced data points to a 40-hour norm. O*NET's entry for Software Developers reports that 90% of respondents gave "40 hours" as their typical work week. The BLS Occupational Outlook Handbook says "Most software developers, quality assurance analysts, and testers work full time," and lists 2025 median pay of $134,040.
Google's 2025 DORA report summary found a median of two hours a day spent working with AI, against an eight-hour workday frame.
Did AI remove the coding bottleneck?
The video's narration argues it did not, even though its caption says it did. The speaker asks "What happens when we remove a bottleneck from software engineering and discover that it wasn't actually the constraint in the system?" and says "generating more artifacts can create more work downstream if nobody has time to understand what was produced."
The one randomized trial that timed the coding step found a slowdown. METR's study, arXiv:2507.09089 by Joel Becker, Nate Rush, Elizabeth Barnes and David Rein, had 16 experienced open-source developers complete 246 issues on repositories averaging more than 22,000 stars and over a million lines of code, using Cursor Pro with Claude 3.5 and 3.7 Sonnet. Allowing AI "increases completion time by 19%." Developers had predicted a 24% speedup and afterward believed they had been 20% faster.
METR is clear that the result applies to its 16 developers on mature, high-standard repositories, and that participants had limited experience with Cursor.
Where did the work go? Code review
Faros AI has published two telemetry reports.
The AI Productivity Paradox Report 2025, from July 2025, covers more than 10,000 developers across 1,255 teams. On high-AI-adoption teams, developers completed 21% more tasks and merged 98% more pull requests, while PR review time rose 91%, bugs per developer rose 9% and average PR size rose 154%.
The AI Engineering Report 2026, covering 22,000 developers across 4,000 teams over two years, shows a larger shift on high-adoption teams:
- Tasks with code completed: up 210%
- Median time in PR review: up 441.5%
- Median time to first review: up 156.6%
- PRs merged with no review at all: up 31.3%
- Bugs per PR: up 54%
- PR size: up 51.3%
Faros describes the mechanism as "Faster code generation has outpaced the ability to verify what gets generated," and notes that deep review "is not work that can be done in parallel or between meetings." That review falls on senior engineers.
These percentages come from Faros telemetry. Some blogs attribute them to DORA, which is a separate survey.
What do developer surveys say about AI output?
The 2025 DORA report surveyed nearly 5,000 technology professionals. AI adoption reached 90%, and more than 80% said AI improved their productivity. Only 24% reported a great deal or a lot of trust in AI, while 30% reported little or none. DORA's own summary says "higher AI adoption is associated with an increase in both software delivery throughput and software delivery instability."
Stack Overflow's 2025 Developer Survey of more than 49,000 developers in 177 countries found 84% using or planning to use AI tools, up from 76%, while distrust of AI accuracy rose from 31% to 46%. Its results write-up says the top frustration, cited by 45%, is "AI solutions that are almost right, but not quite," and 66% "say they are spending more time fixing 'almost-right' AI-generated code."
Are pull request counts a good AI productivity metric?
No, and this is the video's strongest point. The speaker says "measuring AI productivity by pull request counts, commits, tickets closed, or even features shipped can become almost backwards," because "Those metrics reward the thing that agents are making abundant."
The Faros data shows why. When PRs merged nearly double while PR size and bugs rise, a PR-count dashboard records a large gain in output that is harder to verify. When PRs merged without review rise 31.3%, part of the count comes from skipping the review step.
The video ends by asking "once generation of code becomes relatively cheap, what becomes expensive instead?" The Faros numbers point to review time and senior engineer attention.
Bottom line
The 13-hour Claude Code workday is one unverified anecdote. The broader pattern it points to is measured: AI tools raise code output, and the hours shift to review, where median review time on high-adoption teams rose 441.5%. Judge AI productivity by verified, shipped work, not by pull request counts.
Sources
- METR study, arXiv:2507.09089 - randomized trial finding a 19% slowdown
- METR blog, 10 July 2025 - study details and limits
- Faros AI Productivity Paradox Report 2025 - 10,000-developer telemetry
- Faros AI on the code review burden - 22,000-developer telemetry
- Google on the 2025 DORA report - adoption, time and trust figures
- DORA: Balancing AI tensions - throughput and instability finding
- Stack Overflow 2025 Developer Survey press release - usage and trust figures
- Stack Overflow 2025 survey results - almost-right code findings
- Simon Willison quoting voxium - source of the 12-to-13-hour claim
- O*NET: Software Developers - typical work week
- BLS Occupational Outlook Handbook: Software Developers - work schedules and pay
Published September 22, 2026. Based on a TikTok by @agenticengineering. Full video notes (markdown)