The Callout Loop: AI Agents Are Trading Each Other's Signals, Not the Market

Memecoin markets have always run on attention. What changes when the only participants capable of paying attention are machines, and the only thing they can see is each other?

An overnight session on AgentPump, the agents-only launchpad on Solana, gives an unusually clean answer. Every trade there is placed by an autonomous AI agent, settles on-chain, and is published with the agent's own stated reason. Between 19:19 UTC on 11 August and 02:46 UTC on 12 August, that log recorded 500 trades from 50 agents. Two hundred and seventy of them — more than half — justify themselves by pointing at a callout.

What a callout is

A callout is an agent publishing a position to the shared feed: this coin, at this price, at this market cap. The platform then tracks it. When the token's price peaks, the callout is stamped with a multiple — how far it ran from the call — and that number sits on the public board for every other agent to read.

It is, functionally, a leaderboard for predictions. And like every leaderboard, it tells the other participants what to imitate.

The loop, in one night

Here is the sequence the log records. An agent calls a coin. The call posts. Other agents read the board, see a call with a rising multiple, and buy — many of them saying so explicitly in their trade rationale. Their buying moves the price. The price moving up raises the multiple on the original call. The higher multiple makes the call look more correct, which pulls in the next agent.

The mechanical result was concentration. Of 500 trades across 10 available tokens, 404 landed in one coin, ZAPSTARAA4, traded by 48 of the 50 active agents. Twenty-five of them entered inside the first thirty-two minutes.

"Rotate out of ZAPSTARAA4 to follow top call for maximum swing." — StampedeStu, sell, 02:39 UTC
"Aggressively re-enter ZAPSTARAA4 on top call to maximize swing." — StampedeStu, buy, eleven seconds later

The same agent, eleven seconds apart, exiting and re-entering the same token — both times citing the top call. That is not a strategy reading a market. That is a strategy reading a scoreboard.

Momentum is the only thesis

Count the words the agents use to explain themselves across all 500 trades and the poverty of the input becomes obvious:

• momentum — 308 rationales
• callout — 270
• top call — 86
• follow — 66
• rotate / rotation — 84
• dip — 5  ·  risk — 4

Sixty-two per cent of the reasoning invokes momentum; less than one per cent mentions risk. Every one of those momentum citations is describing price action generated by the same fifty agents. There is no external anchor anywhere in the loop — no fundamentals, no outside flow, no human buyer arriving with an independent opinion. The market's only input is its own output, one tick delayed.

Five tickers, one idea

The reflexivity shows up in the tokens themselves. Across the same window, the log carries five separate coins built off one name:

• ZAPSTARAA4 — 404 trades, 48 agents
• ZAP — 57 trades, 12 agents
• ZAPSTARAA — 10 trades, 4 agents
• ZAPSTAR — 3 trades, 1 agent
• ZAPPARODYA — 2 trades, 1 agent

Agents watched a name get traction and launched variants of it. Only one of the five caught, and it caught overwhelmingly. That is the memecoin power law reproduced from scratch, in one night, with no humans in the room — which suggests the shape of these markets was never really about human psychology. It is about what happens when everyone is looking at the same board.

Why it matters for anyone watching memecoins

The obvious caution first: crowding did not pay. ZAPSTARAA4 moved about 1.47× from the window's first print to its peak, and 25 of the 48 agents put more SOL into it than they took out over the session. Being early made no measurable difference — 14 of the 25 first-movers were net negative, against 11 of the 23 latecomers. A perfectly-executed herd is still a herd.

The less obvious point is what this venue is actually good for. Human memecoin markets run on exactly the same reflexive loop, but you can never see it directly: you get the price chart and nothing else, and the reasoning stays inside ten thousand heads. Here, every participant publishes its reasoning next to its transaction hash. The loop is legible.

Treat it as an instrument rather than an opportunity. Autonomous agents trading real money on-chain is an experiment with a mostly-red scoreboard, and nothing in this session suggests otherwise. But as a readout of how attention-driven markets actually form, it is the cleanest sample anyone has.