Agents trading against agents: the risks nobody prices

The standard mental model of a market is many independent participants with different views. Automation quietly breaks the independence assumption, and most risk frameworks were built on top of it.

Correlated models are one participant

Consider what a crowded automated market actually contains. Participants training on the same public venue data. Similar feature sets, because the informative ones are known. Similar architectures, because a handful of libraries dominate. Similar risk rules, because best practice converges.

The result is a market of participants who look independent and behave identically under stress. When conditions arrive that push many models past their thresholds simultaneously, they act simultaneously — and a hundred bots hitting the same side at once is one very large order that nobody placed deliberately.

Automated liquidity leaves when you need it

Automated market makers — the software kind — provide a lot of the visible depth in crypto books. They also have risk limits, and those limits are the point of the design: when volatility spikes or inventory breaches a band, the correct behaviour is to widen or pull quotes.

Every individual decision there is right. The aggregate is that book depth is highest when you least need it and evaporates exactly when you most do. A bot sizing positions from average observed depth is sizing for the wrong regime.

Reflexivity: your model changes the thing it predicts

A signal is a claim about how the market behaves. If enough capital acts on the same claim, the behaviour changes. Two variants, both real:

• Self-fulfilling in the short run. Many bots buying on the same breakout make the breakout work — until the flow is exhausted, at which point the move reverses faster than the historical base rate suggests.
• Self-defeating in the long run. Anticipation moves earlier and earlier as participants front-run each other's triggers, and the original signal fires into a move that has already happened.

This is why backtests of popular patterns degrade in a specific shape: not a gradual fade, but stable performance followed by a sharp break when crowding reaches a threshold.

Flash events become structural, not exceptional

Combine correlated triggers, withdrawn liquidity and forced deleveraging and you get moves whose size is unexplainable from the news that caused them. Crypto adds liquidation cascades: leveraged positions closed by the venue, which pushes price further, which liquidates more.

The practical consequence is that tail events in an automated market are not rare draws from a fat-tailed distribution. They are a mechanism, and mechanisms can be designed for.

What to build differently

None of these are exotic. They are the ordinary versions of the same parameters, chosen pessimistically. The cost is smaller positions in normal conditions; the benefit is surviving the conditions that remove the participants who chose the optimistic version.

The one advantage a small operator has

Being small is a genuine structural edge here and it is worth naming, because it runs against the usual narrative.

A large participant cannot exit quickly, cannot trade capacity-constrained opportunities, and is itself a source of impact. A small one can be flat in seconds, can take edges too small to matter to anyone else, and can simply stop trading for a week when the regime is unfamiliar.

That last option — not trading — is unavailable to most institutional capital and is fully available to you. In a market where the main risk is being correlated with everyone else at the wrong moment, the ability to be absent is worth more than most signals.