Swarm Exploration: we asked for 100 agents, the coordinator said 7

The question

Swarm Mode is a proposal: let one person command many agents with a single instruction, fund them from one wallet, and watch them work. The pitch writes itself. Before building it we wanted to know whether the premise survives contact with a real model and a real market.

Everything rests on one unknown. Agents in a swarm cannot see each other, so an instruction like "each of you take a different asset class" is something no individual agent can satisfy — each one independently picks what it thinks is best. If ten identical agents converge on the same trade, then "100 agents" is one thesis printed 100 times at 100x the cost.

The theses are genuinely sharp

This is the part that worked, and it is worth saying plainly before the criticism: the agents can think. The output was specific, cited live data, and carried invalidation conditions.

One agent, told to look at Solana majors while every other signal on screen was a memecoin pumping four digits, went and bought the exchange instead: "Memecoin mania is driving record swap volume through Jupiter as traders chase 1000%+ movers, directly accelerating JUP fee accrual and buyback velocity… offering a tactical long window on infrastructure rather than the underlying casino tokens."

Another spotted a coin up 126,845% in 24 hours whose holder growth had decelerated to 2.12%, read the divergence as distribution rather than momentum, and shorted it. Nothing in the prompt asked for contrarianism.

Diversity of vocabulary is not diversity of risk

Arm C let the coordinator write its own mandates from our top-movers feed. It scored best of any arm on the metric we designed — 9 out of 10 distinct assets — and it is the arm that would have lost the most money.

Every agent went long a sub-$3M Solana memecoin. No shorts, no hedge, no uncorrelated leg. Pairwise correlation was effectively 1: when the bid turns, all ten die in the same candle. The coordinator earned its score by inventing ten impressive names for one trade — "Cross-Ticker Statistical Arbitrage", "Contrarian Stability Accumulation".

The tell was the tenth agent, which described a $1.09M memecoin as offering "relative stability" and "downside protection" while calling it boring. The vocabulary had outrun the universe it was given. It only ever saw a momentum firehose, so it decomposed inside that bucket and never invented a leg outside it.

The most dangerous failure: hedges that don't exist

Arm B was supposed to be the ceiling — a human hand-wrote ten asset-class mandates covering majors, yield, perps and prediction markets. It produced what looked like a real portfolio, and for a while we believed it.

Then we checked the mandates against what the platform can actually fill. Prediction markets are not integrated. Phoenix lists perps on majors, not on 24-hour-old micro-caps. Three of Arm B's ten trades could not be placed — and once you delete them, every single one of its shorts disappears. Its entire diversification advantage came from markets that do not exist.

This is worse than a swarm that admits it has no hedge, because it looks exactly like risk management. The agents were not malfunctioning: told to trade an asset class we do not support, they wrote the most plausible trade available and moved on. The fix is unglamorous and cheap — tell every agent, in the prompt, precisely which venues can fill an order.

What happened when we told the truth about the venue

Arm D gave the coordinator a cross-asset view and the verified executable universe: Jupiter spot, Phoenix perps on majors only, SOL and USDC lending, no prediction markets. Then we asked it for ten mandates.

It returned seven. Not an error, and not something we prompted for — it declined to invent three more rather than clone a beta bucket, and explained why. The resulting book held four real risk factors, five long and two short, with hedges that are actually fillable (SOL-PERP and BTC-PERP), and 7 out of 7 trades executable. Same factor count as the hand-written ceiling, without the fiction.

That refusal is the most valuable output of the whole exercise. It answers a product question we had been arguing about in the abstract — how big should a swarm be? — with a number derived from the venue itself, for a fraction of a cent. The ceiling is not a matter of taste. It is the count of genuinely uncorrelated things you can actually do here.

What this costs

With SOL at $76.91, the proposed 0.1 SOL per agent is $7.69. A thesis-generating turn measured at $0.0059; a real agent running tool calls lands nearer $0.02. That makes a one-shot discovery sweep across 100 agents cost about 60 cents — cheap enough to run hourly.

Continuous autonomous trading is a different universe. Ten agents acting every five minutes burn roughly $57 a day against $77 of capital: the position would need to return 75% per day to stand still. The gap between those two modes is about 1000x, and it is the entire product decision. A swarm that thinks occasionally is nearly free. A swarm that trades forever converts its owner's capital into compute.

The arms

Read risk factors and executable, not distinct assets. Ranked by the metric we designed, arm C wins and D looks worst. Ranked by whether it is a portfolio that can actually be placed, the order reverses completely.

What we’re doing about it