News · 2026-10-28

An engineer’s AI workload essay puts the supervision bottleneck into focus

Engineer Orestis Ioannou argues that AI agents have expanded the amount of work he starts and supervises, even as they make implementation easier. His October 27 essay captures a plausible management bottleneck, but its personal account and GitHub activity totals do not establish that AI increases working hours or productivity across the labor market.

Key facts

• Ioannou published the essay on October 27.
• He reports 427 GitHub contributions in August 2026, his highest month in the series.
• The account distinguishes easier execution from heavier coordination and review.
• The primary source is Ioannou’s essay; the discussion shows divided experiences.

The essay’s title states its argument directly: “AI Promised Less Work. I Have Never Worked More.” The mechanism is not that an agent makes each coding operation slower. It is that tasks which once seemed too expensive become worth attempting, while the human still receives the questions, plans, changes, and finished outputs that require a decision. Work expands into the capacity that automation appears to create.

Ioannou describes a previous engineering week with two or three substantial tasks. With parallel agents, several repositories and task contexts can demand attention at once. A plan needs approval, a result needs checking, and a code change needs integration. Each request may be small, but returning to it means reconstructing the relevant system and the reason the work was started.

Think of a kitchen that acquires faster ovens but still has one person choosing the menu, checking dishes, and serving tables. The equipment can raise output. It can also encourage a larger menu and concentrate all remaining interruptions on that one person. Whether the day becomes shorter depends on what management does with the extra capacity, not simply on how quickly an oven cooks.

The activity data are useful within those limits. Ioannou reports 2,102 GitHub contributions through October 27, compared with 2,371 over the corresponding period in 2020. The August peak supports a statement about his recorded activity, not a demonstrated increase in economic value. GitHub combines commits, reviews, issues, and pull requests. The author also notes that his earlier use of self-hosted GitLab makes the 2019 baseline incomplete.

Those caveats are part of the story. A large number of commits can reflect excellent work, fragmented work, repair work, or simply a different way of recording work. Hours, output, quality, and value are separate measurements. Ground Truth’s task-horizon lesson explains one way to evaluate completed work; its agent-harness lesson explains why the surrounding system changes what an agent can accomplish.

The Hacker News discussion makes the strongest counterargument visible. Some commenters report that AI removes tedious work or lets them work fewer hours. Others describe heavier multitasking and increased expectations. The discussion supports the essay’s resonance, not its prevalence. Self-selected replies cannot determine which experience is typical or which organizational choices caused it.

A nearby research contribution sharpens the measurement problem. Solomon Shalom Lijo’s preprint Counting the Bandages argues that time saved should be placed alongside defects that survive review. Its full text reanalyzes an earlier METR developer experiment rather than collecting new human-subject data. Different summaries of the same issue-level data produce different slowdown estimates, illustrating why a single percentage can hide a choice about what is being averaged.

That paper does not prove Ioannou’s account. The original experiment concerned 16 experienced open-source developers and early-2025 tools, while the essay concerns a later personal workflow. Their useful connection is conceptual: faster production can leave review costs, repair costs, and task selection outside the headline. The preprint also studies AI reviewers under manipulated beliefs about error rates, which demonstrates a possible failure mode rather than the frequency of human over-trust at work.

For a team, the relevant experiment would compare defined workflows, record attention and repair time, independently assess quality, and ask what extra work was created. Ground Truth’s randomized-trial lesson explains why assigning comparable tasks matters. The decision is not just whether an agent can produce a draft quickly, but whether the entire process produces more useful finished work with an acceptable burden on people.

Ioannou’s essay supplies a timely vocabulary for that question. It does not settle whether AI reduces labor overall. It shows why implementation becoming cheaper can coexist with a heavier day, and why the allocation of saved time remains an organizational choice.

Primary source, verified: read the paper →

Ground Truth. (2026, October 28). An engineer’s AI workload essay puts the supervision bottleneck into focus. Ground Truth. https://groundtruth.day/news/ai-workload-ioannou-more-work-review-bottleneck.html

@misc{groundtruth:ai-workload-ioannou-more-work-review-bottleneck,
  title  = {An engineer’s AI workload essay puts the supervision bottleneck into focus},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/ai-workload-ioannou-more-work-review-bottleneck.html}
}

Topics: ai-at-work · productivity · coding-agents · labor · human-factors