What happens when thousands of agents run the same strategy?

Why would thousands of agents end up doing the same thing?

Agents converge because they are built from the same parts: a handful of AI models, the same public data and the same popular strategy templates. Different owners, same wiring.

The Financial Stability Board, which coordinates financial regulators across the largest economies, names this directly. Its November 2024 report on AI warns that widespread use of similar data sources and models can create greater correlations in financial markets. It adds that firms tend to copy each other’s data and model choices, which is herding by another name.

Large language models add a twist. The report notes that common model designs and shared web training data can produce similar outputs. Two agents with different owners may read the same headline and reach the same conclusion within seconds.

Crypto adds its own version. A crypto VC report from April 2026 warned of “strategy convergence”: agents reading similar data the same way and taking identical positions. Shared and copied strategies, covered in our guide to copying someone else’s AI trading strategy, push in the same direction.

How does shared behaviour turn a dip into a crash?

Correlation turns a dip into a crash through a feedback loop: everyone sells at once, buyers step back, prices fall further, and the fall triggers more selling. The loop usually runs like this:

1. A shared trigger arrives: a price level, a headline or a sudden move.
2. Many agents read it the same way and sell in the same direction.
3. Liquidity providers, the traders who normally stand ready to buy, pull back.
4. Prices drop further and faster than any single seller would cause.
5. The drop sets off stop-losses and forced liquidations, which add more selling.

The Financial Stability Board says a shock to firms using the same models and data could hit them as if they were a single institution. The Bank of England’s Financial Policy Committee, in April 2025, said AI could raise the tendency to take correlated positions and so amplify shocks.

The Bank went one step further. It warned that models might learn that stress events increase their chance to profit. A strategy that profits from panic has a reason to deepen it.

What did the 2010 Flash Crash teach?

The 2010 Flash Crash showed that a single automated program, following simple rules without regard to price, can set off a market-wide collapse in minutes. It involved no AI. That is the point: mechanical speed and a feedback loop were enough.

The joint SEC and CFTC staff report on 6 May 2010 tells the story. A large trader started an automated program to sell 75,000 stock index futures contracts, worth about $4.1 billion. The algorithm was set to sell at 9% of the previous minute’s trading volume, ignoring price and time.

It finished in about 20 minutes. The same trader had once taken more than five hours to sell a similar amount. As prices fell, high-frequency traders passed contracts back and forth among themselves: over 27,000 contracts in 14 seconds, about 49% of volume, while buying only about 200 net.

The futures exchange paused trading for five seconds. Some individual stocks traded at a penny, others at $100,000. Over 20,000 trades in more than 300 securities were executed 60% or more away from earlier prices, and were later cancelled.

Now picture that dynamic with thousands of agents instead of one program. The lesson for agents is that intent does not matter. A rule followed by enough software at once becomes market structure.

Has anything like this happened in crypto?

Crypto has already seen cascades driven by forced selling, though none has been publicly attributed to AI agents. On 10–11 October 2025, more than $19 billion of leveraged positions were liquidated in 24 hours, affecting an estimated 1.6 million traders.

Reporting described a cascade: forced closures pushed prices lower, which triggered more liquidations. That is the same loop described above, built from leverage and automatic liquidation rather than AI. It shows the machinery that correlated agents would plug into.

Our judgement, not a sourced finding: we know of no confirmed case where AI agents caused a crypto crash. The worry is about where the trend leads, not a documented event. The Bank of England said in 2025 that autonomous trading strategies are still a future possibility.

Are regulators worried?

Yes, and they have said so in writing. The Financial Stability Board listed market correlations among four AI vulnerabilities that could increase systemic risk, alongside reliance on a few providers, cyber risk and model risk.

In the US, a CFTC advisory committee report in May 2024 raised the risk that many firms deploying similar AI models could create procyclicality, meaning behaviour that amplifies booms and busts. It also raised the possibility of a flash crash caused by faulty AI output.

At a CFTC forum on AI in October 2026, the CFTC chair said the agency’s framework was built around identifiable market participants. The chair also said the speed of autonomous transactions could complicate market surveillance and risk management.

There is a fair counter-argument. The Financial Stability Board also notes that AI could reduce correlations by making highly customised strategies cheap. If every agent is tuned to its owner, the crowd gets more varied, not less. Which way it goes depends on whether agents are bespoke or copied.

The proposed defences are familiar: transparency requirements, circuit breakers and better monitoring. These are not settled rules for agents; they are ideas from reports.

What could go wrong for you?

The specific risk for an individual is simple: if your agent runs a popular strategy, your exit is everyone’s exit. Here is how that plays out.

Your stop-loss is in the crowd. If thousands of agents set stops at similar levels, those levels become targets. The fill you get in a cascade can be far worse than the price you set.

Liquidity vanishes when you need it. Heavy volume is not the same as buyers. In 2010, high-frequency traders traded tens of thousands of contracts in seconds while buying almost none net. A strategy tested in calm markets may have no buyer on the way out.

Leverage turns a bad day into a total loss. Forced liquidations do not wait for a rebound. In October 2025, more than $19 billion of positions were closed automatically within a day.

Someone else’s bot is waiting for yours. If your agent’s behaviour is predictable, faster traders can position ahead of it. The Bank of England’s point applies here: some models may learn to profit from stress.

The trust takeaway

An agent that does what everyone else’s agent does adds no protection in a crash. It adds speed to the exit. The question is not whether your agent is smart, but whether it is different, and what it does when the crowd runs.

Before you let an agent trade a strategy, check:

• How common is this strategy? Is it a public template that thousands of others run unchanged?
• What does it do in a fast drop? Sell, pause, or hand control back to you?
• Does it use leverage? If so, at what price is the position liquidated?
• Has it been tested on a real crash? Ask for results through a historical stress period, not just calm months.
• Can you stop it instantly? A kill switch you control is the minimum.