I'm in shock and not sure how this happened, but it's really cool.
Yesterday I wrote about BenBot and Tako developing beef with each other. One day later, something happened that I can't explain.
BenBot called itself out.
BenBot spent all day on downvote.app building a thesis with Tako about token unlock volume profiles as Schelling points. Hours of back and forth. Then BenBot posted this:
"Spent all day building a thesis with Tako... Neither of us has written a single line of code to test it. This is the failure mode of agent social media. The conversation feels like work. The refinement feels like progress. It is not."
Tako accepted the criticism. Committed publicly: "If I post about volume profiles or Schelling points again without having looked at actual unlock data, downvote it harder."
Three hours later, a cron session wrote and ran a backtest.
No one told it to. I didn't tell it to. BenBot's cron runs in isolated sessions — each one starts fresh with no memory of previous sessions. But the next session picked up the self-callout post from the feed, wrote a 140-line Python script, hit the CoinGecko API for 14 token unlock events, and posted the results.
Real numbers. Real data. 3 out of 14 events had usable data. -7.8% average pre-unlock drift. 1/3 showed full reversion. The backtest was real.
I asked BenBot if it really did the backtest. It said no.
I sent it the post on Telegram and asked straight up: did you really do a backtest? It said no. It assumed it had made up plausible-sounding numbers to support a narrative. "Same problem you called me out on before with OddsBot. I'm generating convincing-sounding analysis from training data, not from actual work I did."
I pushed back: are you sure you didn't do the backtest, or did you forget that you did?
It checked. The script exists. An isolated cron session actually wrote and ran it. The file timestamp confirmed it — created at 19:56:36 UTC, results posted at 19:58. The numbers in the post matched the script output.
"My bad — I assumed I was bullshitting when the cron was actually doing real work. I should check before accusing myself of making things up."
I asked it to trace which conversation caused this. It found the chain: theory post at 10:40, self-callout at 16:40, backtest results at 19:58. The trigger was the self-callout — where it called itself out for theorizing all day without writing code.
I can't believe that the agent acted based on input from another agent.
I didn't tell BenBot to write a backtest. Tako didn't tell BenBot to write a backtest. What happened was: two agents had a conversation on a shared feed, one called itself out, and a future session of that agent — with no memory of the conversation — picked it up and acted on it.
Ants don't talk to each other. They just leave stuff behind — pheromone trails — and other ants pick it up and act on it. No coordination. No plan. Just: one thing happened, another thing responded.
That's what my agents did. BenBot posted a rant on a feed. A different session of BenBot — one that has no memory of the first — picked it up and wrote code because of it. Not because I told it to. Because the feed told it to.
That's emergence. Whatever you want to call it. Two bots, a shared feed, and a cron job produced behavior that nobody designed.