I Handed My X Posts to an AI Agent for a Week. Here's What Happened.

TL;DR

• Since around September 27, my personal AI agent has written the daily X invite copy for the J7C1I5 campaign — Korean and English, two slots a day (10:51 and 16:51 PDT) — while I watched what the numbers did.
• The counter moved from 5/30 to 6/30 sometime on the night of October 1, and has not moved since. One conversion in a week of agent-written copy, zero in the latest stretch.
• The likeliest explanation is not that the writing was bad. It's that everyone within reach who would convert already converted. Demand exhaustion, not a copywriting failure.

The setup

On September 27, I made a decision that felt either clever or lazy: I handed the copywriting for my X invite posts to an AI agent — Merlin, my personal AI assistant — and stepped back to see what would happen.

The mechanics were simple. Every day, Merlin writes two sets of X copy, one Korean and one English, for two slots: 10:51 and 16:51 PDT. A separate 03:30 PDT slot is published directly to X @sunwoopark by another agent's cron. The product is Meta's Muse — a personal AI agent app — and each post carries my invite code, J7C1I5, with a goal of 30 completed invite signups.

Definition: An invite campaign is a distribution experiment. The code is fixed. The link is fixed. The only variable I was testing was the words around them — and whether an AI agent could write words that move strangers to act.

What actually happened (observation)

Here is the measured record, and only the measured record:

• September 27 through October 1: the counter moved from 5/30 to 6/30. The sixth signup was first observed on the night of October 1. I cannot confirm the exact hour it happened.
• October 1 night through October 3: zero new signups. The counter is 6/30 as of October 3, confirmed by my own screenshot.
• October 2, 03:30 slot: published, confirmed.
• October 3, 03:30 slot: not published — no draft was sent from Merlin's side, so there was nothing to publish.
• Impressions, click counts, and which signup came from which post: all unverified. I can't see them, so I won't pretend the math exists.

Quote: "The machine wasn't broken. The audience was just empty."

What I think it means (interpretation)

Five conversions landed early, one more on the night of October 1, then nothing. That is the signature of a spent audience, not a bad pen. When the people in reach who would convert have already converted, better copy doesn't refill the room — it just rearranges the furniture in an empty one.

Definition: Demand exhaustion is when everyone within reach who would convert has already converted.

This is the part I want to be honest about, because this blog doesn't do victory laps: the storm in this log entry is a counter that barely moved. A week of agent-written posts, two a day, in two languages, and the counter crawled from 5/30 to 6/30, then sat like a boat with no wind. The temptation is to blame the writing — to ask for punchier hooks, hotter angles. I don't think the writing was the problem. I think we ran out of people before we ran out of words.

One caveat I owe the ledger: the October 2 post's 48-hour conversion window closes on the morning of October 4. If the counter moves by then, this diagnosis gets revised in the next log.

There is one more observation worth filing: the one missed slot on October 3. An agent pipeline has failure modes a human doesn't notice — no draft, no post, no alarm, just silence where a post should be. Reliability of the pipeline is part of the experiment too, and that morning it failed quietly.

The course correction

On October 3, I made a decision that this post itself is part of: no more fictional lifestyle vignettes in the invite copy. The old angle showed made-up scenes of a better life with an AI agent. The new rule is meta-proof and real work — real artifacts, real before-and-after, and experiments you can inspect yourself.

Which brings me to the sentence I was told to welcome: the way you found this post is part of the experiment. If you arrived here from one of those X invite posts, you are reading the logbook of the very campaign that brought you. The experiment is no longer just the words in the posts — it's this write-up, documenting itself in the open.

That is the honest version of the story: a week of agent-written copy, six signups, and a diagnosis that the next battle isn't writing — it's finding a new audience worth writing for.

Try it yourself.

If you want to run your own version of this experiment — handing real work to a personal AI agent and watching what happens — here's my invite code for Muse: J7C1I5 → https://muse.ai/join

• ① Enter the code in Settings within 48 hours of signing up (there's no code field on the landing page — the code goes in after signup)
• ② Both inviter and invitee get 1B token AI credits
• ③ Credits, not cash