Concepts
Meat proxy
A meat proxy, as named in the room, is a person who forwards AI-generated text without reading, understanding, or validating it — a relay between AI systems. The examples were mundane and everywhere: an error in Slack, pasted into the model, pasted into the terminal; a prod error, same path; a CI failure, same path.
Last year that was two hours around five or six hours of high-IQ work, and it was tolerable. Now the copy-paste is eight or nine hours. The human did not get slower. The agents got faster, and the paste layer did not go away.
Why it matters
Section titled “Why it matters”Human-in-the-loop sounds like governance. At this volume it is how you fail to get the return.
From the room
Section titled “From the room”Last year the day was filled with five or six hours of high-IQ work, and the two hours of copy-pasting were tolerable. Now it’s no longer tolerable, because it became eight or nine hours of copy-paste. And so we’re turning into these meat proxies.
The first physical version of the same problem: agents that only run on a laptop stop when the laptop closes. Move the work into the cloud, wake an agent from a webhook, and the human’s remaining job is policy — including “a person still approves what ships to prod” — not being the pipe.
A 50,000-line PR from one developer is the review-queue version. The answer given in the room was not less AI. It was more AI on the steps that were drowning the humans.
Where founders get it wrong
Section titled “Where founders get it wrong”- Keeping coding agents on a laptop that sleeps when you go to a meeting.
- Requiring a human paste for every CI failure, prod alert, or GitHub issue.
- Calling “human in the loop” a control when it is actually the bottleneck.
- Answering a 50,000-line PR with less AI instead of putting agents on review, test, and operate.
Go deeper
Section titled “Go deeper”- Why AI isn’t making engineering 10x faster is the playbook this sits inside.
- Related concepts: Amdahl’s law, AI-native.