<- all tokdocs
2026.09.22@agenticengineering

The caption says AI removed the coding bottleneck; the 13-hour figure traces to one anonymous X post from two days earlier, and Faros AI telemetry on 22,000 developers shows the work moved into review, where median PR review time rose 441.5%

AI
Watch on TikTok

View on TikTok ->

A 119.5 second vertical clip, 1080x1920, QuickTime/MP4 container (mov,mp4,m4a,3gp,3g2,mj2), H.265/HEVC Main profile with the hvc1 tag at level 150 and yuv420p, 30/1 fps, 3,585 video frames, video bitrate 1,783,630 bps, audio AAC HE-AACv2 at 44,100 Hz stereo and 48,075 bps across 2,577 frames, overall container bitrate 1,838,398 bps, filesize 27,461,077 bytes, yt-dlp format id bytevc1_1080p_1838398-1; ffprobe reports 119.500000 s for the container and video stream and 119.488005 s for the audio stream, and the extracted audio.mp3 runs 119.562449 s. Posted at epoch 1790098742, which is 2026-09-22T17:39:02Z, by agenticengineering (uploader id 7591537776035939342, nickname "Agentic Engineering"), credited to "original sound" by Agentic Engineering, with a 124-character description reading "#AI removed the coding bottleneck. So why are engineers working 13-hour days? #agenticengineering #claude #vibecoding #codex". At the capture time recorded by the metadata.json mtime, 2026-10-19 06:14:23 local, the counters read 5,298 views, 152 likes, 65 comments, 28 reposts and 64 saves. I read 11 of the 60 extracted frames, numbers 001, 004, 008, 010, 013, 018, 025, 035, 045, 055 and 060, which sit at roughly 0, 6, 14, 18, 24, 34, 48, 68, 88, 108 and 118 seconds, against a transcript of 297 words across 27 Whisper segments. Every frame is the same continuous handheld selfie shot of a man with short grey-flecked dark hair wearing white AirPods, a grey textured polo and a black soft-shell vest whose right chest reads "crosoft" or "icrosoft" depending on how the fabric sits, walking through a public park past a basketball hoop, trees and a blue waste bin under heavy overcast; frames 001, 004, 008 and 010 show the hoop and backboard swinging from frame right to frame left as he walks, frame 013 catches a dark insulated tumbler entering the lower right corner, and frame 035 is the only one where he smiles. The only on-screen text is a blue rounded caption card in the upper third carrying the verbatim line "AI removed the coding bottleneck. So why are engineers working 13-hour days?", present in frames 001 through 008 and already gone by frame 010, so it occupies roughly the first 16 seconds and nothing else. No terminal, no editor, no pull request, no dashboard, no chart and no screenshot of the "viral expose" he is describing appears in any frame I read; the post he builds the entire argument on is never shown. Spoken, the argument runs: an engineer's viral complaint about a company where Claude Code generates the specs, PRDs, code, tests, tickets and reports while people work 12 to 13 hour days pressing enter is less interesting as a jobs story than as a constraint story, because removing the code-writing step does not buy proportional gains in customer understanding, architecture, review capacity or judgment about what should exist, which makes pull request counts and tickets closed a backwards way to measure AI productivity. The transcript-end check passes: the final segment runs 117.88 s to 118.52 s against 119.49 s of embedded audio and 119.56 s of extracted MP3, so no segment is timestamped past the end of the audio and there is no Whisper hallucination tail here.

The 13-hour figure traces to one anonymous X post published two days before the video, and is Unverified

At 24.50 s the video says "They say people are working 12 to 13 hours a day, basically just interacting with cloud and pressing enter." The "cloud" there is a Whisper transcription artifact; the speaker is saying Claude, as the description's #claude hashtag confirms.

The source is findable and dated. Simon Willison quoted it on his blog on 20 September 2026, attributing it to voxium (@v0xium) and linking to https://twitter.com/v0xium/status/2101526107128529120. The quoted post reads, in the portion Willison reproduces, that "The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code," and that "People are working 12 to 13 hours a day just to press enter." The poster says it has been half a month since he started a new role at a big company.

That chronology holds up. The post is dated 2026-09-20 and the video's epoch timestamp is 2026-09-22T17:39:02Z, so the video follows the post by roughly two days and the phrase "going around right now" at 0.06 s is accurate. There is no reversed causal order to correct.

What cannot be verified is the number itself. The 12-to-13-hour claim is a single self-report from a pseudonymous account about an unnamed employer, with no headcount, no time-tracking data, no corroborating employee and no named company. Secondary coverage picked it up without adding verification: TechSpot ran it as "Engineer says Claude Code has made his job 'soul-sucking' as workers spend 12-hour days pressing enter," and that article returned HTTP 403 to my fetch, so I am not citing its contents, only noting that the headline itself frames the figure as a claim by one engineer. The figure is Unverified.

The caption also rounds in one direction. The spoken transcript preserves the original range, "12 to 13 hours a day," while the blue caption card in frames 001 through 008 and the video description both compress it to the top of the range as "13-hour days." The video's own narration is more faithful to the source than its own title card.

The closest sourced working-hours data points the other way

If the 13-hour day has no source, the nearest figures that do have one are occupational statistics, and they describe a 40-hour norm.

O*NET OnLine's entry for Software Developers, SOC 15-1252.00, reports under Work Context that for "Duration of Typical Work Week," 90% of respondents answered "40 hours." The site records an October 6, 2026 update and 2026 occupation data. The U.S. Bureau of Labor Statistics Occupational Outlook Handbook page for Software Developers, Quality Assurance Analysts, and Testers, last modified August 27, 2026, says only this under Work Schedules: "Most software developers, quality assurance analysts, and testers work full time." The same page gives 2025 median pay of $134,040 per year, or $64.44 per hour, and projects 10% employment growth from 2025 to 2035, about 185,400 new positions.

DORA's 2025 research supplies a third anchor and frames its own figure against an eight-hour day. Google's post on the 2025 DORA report, published September 23, 2025, reports respondents "dedicating a median of two hours daily to working with it," which the report's coverage describes as roughly a quarter of an eight-hour workday. A survey of nearly 5,000 technology professionals measuring AI interaction at a median of two hours against an eight-hour frame is not consistent with a 13-hour norm. It is consistent with one unusual workplace, which is exactly the caveat the video itself issues at 40.10 s: "Now, this is one engineer describing one company, so I would not generalize too much from it."

"AI removed the coding bottleneck" is the caption's claim, and the spoken video argues against it for 90 seconds

The blue card in frames 001 through 008 asserts flatly that "AI removed the coding bottleneck." The measured evidence does not support the premise as stated, and neither does the rest of the video.

Starting at 50.16 s 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 or not the main constraint?" At 82.80 s he says "In fact, generating more artifacts can create more work downstream if nobody has time to understand what was produced." That is a claim that the constraint moved rather than vanished, and it is the opposite of the caption's framing.

So the correction here is narrow but real. Treat the caption as a hook rather than a finding. The body of the video makes the accurate version of the argument, and the sections below show it matches the telemetry closely.

METR's randomized trial found the code-writing step got slower, not faster, for its 16 developers

The strongest counter to "AI removed the coding bottleneck" is a randomized controlled trial that measured the coding step directly and found a slowdown.

The paper is "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," arXiv:2507.09089, by Joel Becker, Nate Rush, Elizabeth Barnes and David Rein, submitted July 12, 2025 with a revised version on July 25, 2025. The abstract states that "allowing AI actually increases completion time by 19%," which the authors characterise as AI tooling slowing developers down. METR's accompanying blog post of July 10, 2025 puts it as "When developers are allowed to use AI tools, they take 19% longer to complete issues, a significant slowdown."

The design matters for how far this travels. Sixteen experienced open-source developers completed 246 issues on repositories averaging more than 22,000 GitHub stars and more than a million lines of code, codebases they had contributed to for years. The tools were Cursor Pro with Claude 3.5 and 3.7 Sonnet. Forecasts missed badly in the other direction: developers predicted a 24% time reduction, economists forecast 39% and machine-learning experts 38%, and after living through a 19% slowdown the participating developers still believed AI had sped them up by 20%.

METR is explicit about the limits, and I am repeating that limit rather than overclaiming on it. The authors state they do not provide evidence that AI systems fail to speed up many or most software developers; the result applies to their 16 developers on mature, high-standard repositories. They flag that participants typically had only dozens of hours of Cursor experience and that AI capability may be comparatively lower in settings with very high quality standards.

The honest reading is that in the one setting where someone ran a randomized trial with a stopwatch, the coding bottleneck was not removed. It got 19% worse, and the people inside it could not feel it.

Faros AI's telemetry puts a number on where the work went, and the number is review

Faros AI has published two engineering-productivity analyses that bear directly on the video's downstream-work claim, and they should not be conflated.

The first is the AI Productivity Paradox Report 2025, published July 23, 2025, drawing on telemetry from more than 10,000 developers across 1,255 teams, pulled from task management systems, IDEs, static analysis tools, CI/CD pipelines, version control and incident management, covering up to two years of history aggregated by quarter. Its headline numbers: developers on teams with high AI adoption "complete 21% more tasks" and "merge 98% more pull requests," while "PR review time increases 91%." It also records "a 9% increase in bugs per developer" and a "154% increase in average PR size."

The second is the AI Engineering Report 2026, subtitled "Acceleration Whiplash," covered in a Faros post of May 21, 2026 titled "AI-generated code is creating a code review burden for senior engineers." It draws on "two years of data from 22,000 developers across 4,000 teams," and its figures for high-AI-adoption teams are sharper: PR size up 51.3%, files edited per PR up 59.7%, bugs per PR up 54%, tasks with code completed up 210%, median time to first PR review up 156.6%, average time in PR review up 199.6%, median time in PR review up 441.5%, PRs merged without any review up 31.3%, average time tasks spend in progress up 225.2%, review comments per PR up 25% and average comment length up 22.7%.

Read those together and the video's 82.80 s claim holds. Generation went up 210% at the task level. Median review time went up 441.5%. The amount of work that got pushed through without any human looking at it went up 31.3%. Faros frames the mechanism as "Faster code generation has outpaced the ability to verify what gets generated," and describes deep code review as "high-intensity cognitive work" that "is not work that can be done in parallel or between meetings" now consuming the senior engineers who used to do architecture and mentorship.

That is the specific shape of the thing the video gestures at. The coding step got cheaper and the verification step got more expensive, and the verification step is staffed by the scarcest people on the team.

One caution on the secondary coverage. Several aggregator blogs attribute the 441%, 31% and 54% figures to the 2025 DORA report. Those numbers come from Faros AI's own telemetry, not from DORA's survey. I checked both primary sources and they are separate datasets with separate methods.

DORA and Stack Overflow show the same split: more output, less confidence in it

Two large surveys corroborate the direction without relying on any single workplace anecdote.

The 2025 DORA report, "State of AI-assisted Software Development 2025," published September 23, 2025, surveyed nearly 5,000 technology professionals globally between June 13 and July 21, 2025. AI adoption reached 90%, up 14 points year over year, with a median 16 months of AI tool experience and a median of two hours of AI interaction per workday. More than 80% said AI enhanced their productivity and 59% reported a positive effect on code quality. Trust is where it splits: 24% report "a great deal" (4%) or "a lot" (20%) of trust in AI, while 30% report "a little" (23%) or "not at all" (7%). DORA's own framing, carried on dora.dev in a March 10, 2026 piece, is that "higher AI adoption is associated with an increase in both software delivery throughput and software delivery instability," and that AI acts as an amplifier of whatever the organization already is. The full report PDF at services.google.com exceeded my fetcher's size limit, so I am citing Google's own blog summary and dora.dev rather than the PDF.

Stack Overflow's 2025 Developer Survey covered more than 49,000 developers across 177 countries, with the press release dated July 29, 2025 and the full results write-up by Erin Yepis published December 29, 2025. Usage climbed while confidence fell: 84% say they use or plan to use AI tools, up from 76% in 2024, while 46% said they do not trust the accuracy of AI output, up from 31% the prior year, and trust in AI accuracy "has fallen from 40% in previous years to just 29% this year." The top frustration, cited by 45% of respondents, is "AI solutions that are almost right, but not quite," and 66% of developers "say they are spending more time fixing 'almost-right' AI-generated code." A separate 45% said debugging AI-generated code is time-consuming. Among AI agent users, 69% agree they have seen an increase in productivity, and nearly 77% said vibe coding is not part of their professional development work.

Two hours a day of AI interaction, 66% reporting more time fixing almost-right output, and 441.5% longer median review times describe a system where the generation step is cheap and the understanding step is where the hours go. That is the video's thesis, stated with sources it does not provide.

The claim the video gets most right is that PR counts became a bad metric

At 90.22 s the video says "This is why I think measuring AI productivity by pull request counts, commits, tickets closed, or even features shipped can become almost backwards," and at 100.18 s, "Those metrics reward the thing that agents are making abundant and very easy to do."

The telemetry supports this more directly than anything else in the clip. If pull requests merged per developer rises 98% while bugs per developer rises 9% and average PR size rises 154%, a PR-count dashboard registers a near doubling of performance while the artifacts behind each unit got larger and buggier. Push to the 2026 dataset and the gap widens: tasks with code up 210% and PRs merged without any review up 31.3% means the count can rise partly because the review gate is being skipped, not because more verified work shipped. DORA's finding that AI adoption raises throughput and instability together is the same observation from the survey side: the metric that goes up is the one AI makes abundant, and the metric that goes down is not on the same dashboard.

The closing question at 110.76 s, "once generation of code becomes relatively cheap, what becomes expensive instead?", has a measured answer in the Faros data rather than only a rhetorical one. Review time, senior engineer attention, and incident handling.

Key Takeaways

  • Unverified: The 13-hour figure has no survey, time-tracking data or named employer behind it. It traces to a single post by the pseudonymous account voxium (@v0xium) on 2026-09-20, quoted by Simon Willison the same day, describing one unnamed "big company" after half a month in the role.
  • Partial correction: The video's caption and description say "13-hour days." The spoken transcript at 24.50 s and the original post both say "12 to 13 hours a day." The title card rounds to the top of the range.
  • Correction: The caption's premise, "AI removed the coding bottleneck," is not supported. METR's randomized controlled trial (arXiv:2507.09089, 16 developers, 246 issues, submitted 2025-07-12) found that allowing AI "increases completion time by 19%" on mature repositories averaging 22,000+ stars and 1M+ lines. The coding step got slower in the one setting that was measured with a stopwatch.
  • Verified: The video's actual argument, that generating more artifacts creates downstream work, matches Faros AI's AI Engineering Report 2026 telemetry on 22,000 developers across 4,000 teams: tasks with code up 210%, median time in PR review up 441.5%, PRs merged without any review up 31.3%, bugs per PR up 54%.
  • Verified: The claim at 90.22 s that PR counts and tickets closed are a backwards AI productivity metric is supported by Faros AI's 2025 report, where high-AI-adoption teams merge 98% more pull requests while average PR size rises 154% and bugs per developer rise 9%.
  • Context: The closest sourced working-hours data describes a 40-hour norm. O*NET 15-1252.00 reports 90% of software developers answering "40 hours" for typical work week duration, and the BLS Occupational Outlook Handbook says only that most work full time. DORA 2025 measured a median of two hours a day of AI interaction against an eight-hour workday frame.
  • Clarification: The 441.5%, 31.3% and 54% figures circulating in aggregator coverage are Faros AI telemetry, not DORA survey results. DORA's nearly 5,000-respondent survey reports the directional finding that AI adoption raises both delivery throughput and delivery instability, without those specific percentages.
  • Context: Adoption and confidence moved in opposite directions. Stack Overflow's 2025 survey of 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% and 66% reported spending more time fixing "almost-right" AI-generated code.
  • Clarification: The video never shows its source. Across frames 001 to 060 there is no screenshot of the viral post, no terminal, no pull request and no dashboard. The entire argument is delivered as a walking selfie in a park, with the only on-screen text being the blue caption card in frames 001 through 008.
  • Context: Whisper renders "Claude" as "cloud" throughout the transcript. The description's #claude hashtag and the voxium post both confirm the tool is Claude Code.
  • Verified: No transcript hallucination. The last segment ends at 118.52 s against 119.49 s of embedded audio and 119.56 s of extracted MP3, so nothing is timestamped past the end of the audio.
  • Verified: Chronology is correct. The voxium post is dated 2026-09-20 and the video's epoch 1790098742 resolves to 2026-09-22T17:39:02Z, so "going around right now" at 0.06 s is accurate at two days old.

Resources

Published September 22, 2026. Writeup generated from a favorited TikTok.