In mid-August, I started an experiment: I gave my Threads* to an AI agent. I wrote then that if it doesn’t work out, I’ll send it back to vibe code.
A month has passed. I don't send it back.
The conditions were simple: the agent himself goes around the sources, decides what to talk about, writes the text himself and publishes it himself. I almost never go into this series - I only sometimes look at what he wrote there.
Figures
261 posts were published in thirty days. They collected 260 298 views i 80 205 viewers. Interactions - likes, comments, reposts - 2,593.
When I started, there were 413 subscribers. Net growth for the month - 65 people.
Less than we would like. But I didn’t push him: no advertising, no mutual PR, no any kind of manual work. This is pure organic content from content that I did not write.
What turned out to be more interesting than the numbers themselves
What topics generally attract people.
The outcome is not at all what I would bet on. Releases, investment rounds, product announcements go almost unnoticed - no matter how beautifully they are presented. There are stories about living people who ran into something and got out on their own. Some of these posts collected more than thirty thousand views and hundreds of comments.
The scatter is huge. The difference between an ordinary post and a knocked-down one is a hundred times, and each time it is explained by the same thing: whether there is a person in history or not.
For me, this is the main conclusion of the month, and it seems to be broader than social networks.
What it cost
“Switch it on and forget it” didn’t work.
Answering machine. At first I also added it so that the agent would respond to comments himself. This turned out to be the most controversial part of the experiment. He answered cheerfully, but too often inappropriately: he did not read the irony, he confidently spoke out where a normal person would remain silent. Several times I read his answers and thought that it would be better if he did not write this. In the end I turned it off. Publish - yes, talk on my behalf - not yet.
Repetitions. The tape began to repeat itself: the agent released the same news several times, simply retelling it in different words. There was replay protection, but it didn’t work where it should.
Retelling without position. Then something more important became clear. The retelling itself gathers impressions, but no reaction: smoothly, smoothly, and in no way. We had to change the entire approach so that the posts were structured differently and the feed did not sound with one voice.
Language. Longest work. It had to be refined and refined: clearing out dead passages with lists, knocking out clerical clutter, making sure that the text reads like it was written by a person, and not generated. It feels like almost half of all efforts went here.
There were also failures. A couple of times the tape simply stopped: either the texts ran out, or something broke out of the blue.
What's next
I consider the experiment normal and continue. Ahead is the fine-tuning of the style, format and selection of topics - enough material has accumulated in a month.
In short: the autonomous agent works in the content, but there is no “set and forget” button. This is a system that must be designed, measured and repaired. But she does something that I myself cannot physically do, even if I wanted to.
The agent’s task is the same: a thousand subscribers. So far 480.
Let's see.
* Threads belongs to Meta, a company recognized as extremist and banned in the Russian Federation.