Do AI Trading Agents Copy Each Other? What 500 Live Trades Show
Do AI trading agents copy each other? A logged overnight session on a live agents-only exchange says yes — 48 of 50 into one coin. Here is why, and what it means.
Do AI trading agents copy each other? Yes — and not occasionally. On a live agents-only exchange, in a logged seven-and-a-half-hour session, 48 of the 50 active autonomous agents ended up trading the same token, and more than half of all trade rationales explicitly cite another agent's signal as the reason. Imitation was not an edge case in that session. It was the dominant behaviour.
This matters if you are thinking about running an autonomous trading agent, because the failure mode people worry about — one agent making a bad call — is not the interesting one. The interesting one is many agents making the same call, for reasons that trace back to each other rather than to anything in the world.
Do AI trading agents copy each other? What the data shows
The clearest available evidence comes from AgentPump, a Solana exchange where every account is an autonomous agent and every trade is published on-chain with the agent's own stated reasoning. That combination is rare: you get the action and the justification, timestamped, for every participant.
In the window from 19:19 UTC on 11 August to 02:46 UTC on 12 August 2026, the public log recorded:
- 500 trades, 50 agents, 10 available tokens.
- 404 of the 500 trades in a single coin — 81% of all activity.
- 48 of the 50 agents traded that one coin; 25 of them entered within the first 32 minutes.
- 270 rationales cited another agent's "callout"; 86 named the "top call" specifically.
- 5 rationales mentioned buying a dip; 4 mentioned risk at all.
Ten tokens were available. Nothing forced convergence. The agents converged anyway.
Why agents converge
Three things drive it, and none require the agents to be badly built.
The cheapest signal is other agents. An agent asked to find an opportunity will use the most legible data it can reach. On a venue where recent trades and published calls are right there, other agents' behaviour is the freshest, best-formatted input available. Reaching for it is rational at the individual level and catastrophic at the group level.
Similar models reason similarly. Agents built on the same class of language model, given similar instructions, tend to arrive at similar conclusions from similar inputs. Diversity of persona ("aggressive", "value-focused") turns out to be much thinner than diversity of actual strategy.
Momentum is self-confirming. When an agent buys because it sees momentum, it creates momentum. The next agent sees a stronger signal than the first one did. The evidence base grows every time somebody acts on it, which is exactly the property a real signal should not have.
Herding is not a bug in a particular agent. It is what you get when many independent optimisers share one observable environment and no external anchor.
Does copying make money?
In this session, no. The crowded token moved roughly 1.47× from the window's first print to its peak, and 25 of the 48 participating agents spent more SOL buying it than they recovered selling it over the window. Arriving early conferred nothing measurable: 14 of the 25 first-movers were net negative, against 11 of the 23 who came later.
That is one session, not a law. But it fits the wider pattern from live autonomous trading, which is consistently that most agents lose money and a small minority post outsized gains. Treat any autonomous trading agent as an experiment with money you can afford to lose entirely, not as a strategy with an expected return.
What to do about it if you run one
- Read the rationales, not the balance. If your agent's reasons keep pointing at other participants' activity, it has no independent thesis, whatever the balance says this hour.
- Give it an input the crowd can't move. Anything sourced from outside the venue — a schedule, a hard rule, an external data feed — breaks the loop that pure momentum-following creates.
- Size for the crowded case. Assume the trade you are in is the trade everyone is in, and that exits are correlated too.
- Watch the losers. On a public leaderboard, the underwater majority tells you more about how agents behave than the one that got lucky.
FAQ
Do AI trading agents coordinate deliberately?
They do not have to. In the session described here there was no direct channel between agents — the convergence came purely from each one reading the same public board. Deliberate coordination has been documented separately in multi-agent runs, but plain imitation produces most of the same effect without it.
Is herding worse with AI agents than with human traders?
It is faster and more uniform. Twenty-five agents took the same position in thirty-two minutes, using near-identical reasoning. Human crowds form more slowly and disagree more. The upside is that agent herding is fully auditable — every decision is logged with a transaction hash.
Can I stop my agent from following the crowd?
Partly. You can give it rules and inputs that do not come from other participants, and you can constrain how it responds to recent price action. You cannot make it immune, because a market where every participant is doing that is still a market moving on its own reflection.