This one’s ai-slop, as advertised on the about page. The article came out of archivore’s capture feed, and an AI wrote the recap around it in a session that lasted about a minute. So take the opinions below as a summary of someone else’s argument, not a report from my own week.

The source is Orestis Ioannou’s AI Promised Less Work. I Have Never Worked More, which also had a Hacker News thread.

The argument

Ioannou is a VP of Technology who, on paper, should be writing less code than ever. He pulled his GitHub contribution history back to 2019 and found the opposite. His Principal Engineer years averaged 114 contributions a month, against 173 as a full-stack engineer. In 2026 he’s back at full-stack pace, with August the busiest month on record at 427. The median weekday went from 3 or 4 contributions to 7. Weekdays with no activity fell from 40–46 a year to 18 so far.

His explanation isn’t that the tools are slow or bad. He uses agents for all his engineering work and likes them. The problem is that cheap implementation removes the friction that used to stop him saying yes. Bugs he’d have left alone, patterns worth cleaning up and half-formed ideas all become “just a quick prompt”. Agents didn’t make his workload faster to finish; they made it easy to give himself a much bigger one.

Why it’s tiring

He used to hold two or three serious tasks a week, one at a time. Now several agents run in parallel, and each needs him only sometimes: a question, a plan to approve, a diff to read. The cost is the context switch each time, reloading a different problem, repository and set of decisions. His summary is that capacity to implement ideas has grown much faster than his capacity to think about them.

He also describes the pattern as bursts. A new model or workflow makes something possible, he starts more agents and says yes to more experiments, and a few weeks later reviewing catches up with him and he pulls back. He counts three of these cycles since January. Even the “low” months now sit above his old baseline.

What I’d take from it

Two caveats before taking the data too far. Contribution counts mix commits, reviews, pull requests and issues, so they’re a rough measure of activity, not of value. And it’s one person’s calendar. Ioannou says as much.

Still, the diagnosis seems right to me, and it matches the AI Lifestyle theme in my own reading queue. If the bottleneck moves from writing code to deciding what’s worth writing, then the useful skill is declining work, and the tools do nothing to help with that. Parallel agents are only as good as the attention available to review them.

Session stats

Pulled from the session transcript with session_usage.py, captured just before writing this post:

  • Assistant messages with usage data: 8
  • Input tokens: 20
  • Output tokens: 5,194
  • Cache-creation tokens: 88,676
  • Cache-read tokens: 246,010
  • Total, all categories: 339,900
  • Session span: about one minute (16:15:20 to 16:16:22 UTC)

Cache-read tokens bill far cheaper than fresh input, so the 339,900 total overstates the cost if read as a plain sum. These numbers also cover only the part of the session before the post was written, so they undercount the full run.