The Next Crowded Trade Will Be Built by Machines

We are at a historical confluence of transformational forces that simultaneously drive digitisation, machine intelligence, and automation. There is no better reckoning of this seismic shift than the efforts put into the fast evolution and adoption of AI agents. The practice of investing and trading is no exception: AI agents are commoditising quantitative trading systems that once were only attainable by a few dominating household names. While this rapid democratisation opens doors for many to profit from sophisticated investing, it also raises concerns regarding a lingering risk: crowding.

The concern

Investors and trading firms are racing to put AI agents to work designing and running strategies. The productivity case is obvious. The systemic case is not. If many firms hand similar reasoning tools the same public data and ask them to find alpha, the agents will tend to find the same alpha. That pushes more capital into the same positions at the same time — the oldest systemic-risk problem in markets, but arriving through a new and faster channel.

Why the concern is justified

Crowding is not a theoretical worry. It has already cost institutions real money and is a well documented, priced risk in equity markets.

In August 2007, a group of quant hedge funds running similar factor strategies all hit their risk limits within days of each other. The ensuing unwind erased roughly a decade of alpha in 72 hours.
Since then, academic and industry research has consistently shown that stocks heavily held by sophisticated institutions earn a return premium because they crash harder during drawdowns. The premium is compensation for tail risk, not a free lunch.
Institutional concentration has grown. The top ten asset managers control a meaningful share of US equity holdings. Factor and smart-beta products now hold well over $2 trillion globally.
This concentration is further compounded by leveraged ETF products, which channel speculative retail capital into the same crowded factor exposures.

So the mechanism — many arbitrageurs chasing the same idea, exhausting liquidity, running for the same exit when it turns — is not speculative. It is measured, priced, and cyclical.

What we have learned so far

A few durable lessons sit inside that history:

Crowded trades can be profitable most of the time. That is exactly what makes them dangerous.
Publication accelerates convergence. Once an edge is written up, institutions trade it more aggressively and more uniformly.
The unwind is faster than the buildup. Crowding accumulates over quarters; it resolves in days.
Heterogeneity among human researchers has been a quiet source of market stability.

Likely future scenarios with AI agents in the loop

Scenario 1 — Quiet amplification. AI agents get adopted gradually across mid-tier asset managers. Each firm believes its fine-tuning and private data make its agent unique. In aggregate, because agents draw on the same foundation models and the same public financial corpus, strategies correlate more than expected.

Scenario 2 — Reflexive de-crowding. Agents are explicitly trained or prompted to avoid crowded positions. New, less obvious crowding shows up in places current metrics don't measure.

Scenario 3 — Fragmentation. Firms invest heavily in proprietary data, custom fine-tuning, and differentiated objectives. AI-run strategies end up less correlated than today's quant landscape.

Scenario 1 is likely the base case. Scenario 2 is plausible but limited by the lag in crowding data; Scenario 3 requires sustained investment in differentiation that is beyond the reach of most players.

What does this mean for investors

Sophisticated portfolio managers are already expected to track crowding in their own book and monitor it across peers. For trading firms deploying AI, we should assume the strategies their AI agents discover are also being discovered by someone else's agents. The edge worth having is not a better agent — it is a clearer view of what everyone else's agent is likely to be doing.