Regulators Are Worried About AI Trading Agents. Here Is What They Are Actually Afraid Of
In one week, two of the most powerful financial regulators in the world said the same uncomfortable thing out loud: autonomous AI agents could make markets break faster than humans can react. On November 16, Federal Res
In one week, two of the most powerful financial regulators in the world said the same uncomfortable thing out loud: autonomous AI agents could make markets break faster than humans can react.
On November 16, Federal Reserve Governor Michael Barr told a conference in Milan that AI agents could turn a bank run into an event that unfolds in minutes, not days (Bloomberg, CoinDesk). Four days earlier, Jonathan Hall of the Bank of England's Financial Policy Committee warned that autonomous trading agents could collude, amplify volatility and threaten liquidity (FOW).
The headlines read like science fiction. The concerns underneath them are concrete, and most of them have practical answers. If you use, build or are thinking about renting an AI trading agent, these are the four fears worth understanding.
Fear 1: Speed without a pause button
A human who reads bad news about a bank needs time: to check, to call someone, to log in, to move money. An agent needs none of that. Barr's point was simple. If thousands of agents watch the same signals, such as a rate gap or a whiff of stress, and are allowed to act on them instantly, the run that took days in 2008 and hours in 2023 could take minutes.
The problem is not that agents are fast. Markets have had fast algorithms for decades. The problem is fast action with no circuit breaker in between the decision and the money.
What answers it: limits that live outside the agent. A daily loss limit, a maximum size per trade, an automatic stop after a set drawdown, and a kill switch that a person or a platform can pull. The key detail is where those limits sit. If the agent can rewrite its own limits, they are suggestions. If the platform enforces them above the agent, they are brakes.
Fear 2: Everyone running the same brain
Hall's warning about collusion sounds like agents plotting together. The more realistic version is quieter: thousands of agents built on a handful of similar models, trained on similar data, reading the same news, reach the same conclusion at the same moment. Barr flagged the same risk as herding in wholesale trading.
Nobody has to coordinate anything. When everyone sells because the model said "sell", liquidity disappears exactly when it is needed.
What answers it: diversity and transparency. Strategies that differ in how they decide, not only in their names; and published records that let anyone see whether an "independent" agent is really just following the crowd. An agent whose decisions you cannot inspect is one you cannot tell apart from the herd.
Fear 3: Who actually gave the order?
This was the most striking part of Barr's remarks. Large banks, he said, are already seeing AI agents try to take actions in retail customer accounts without any way to verify that the customer agreed. Several of the largest US banks reportedly block outside agents from customer accounts altogether.
It is the same problem in trading. When an agent places an order, three questions need clear answers: who authorised it, what exactly it was allowed to do, and can it touch money it was never meant to touch?
Barr's proposed answers were stricter identity checks, consent tied to specific transactions, and hard limits on what agents can do.
What answers it: narrow permissions by design. Access that can trade but cannot withdraw. Agents that run in isolation from each other and from your other accounts. A test period on paper or demo money before anything real is at stake. Robinhood, for example, keeps its agents in a separate, dedicated account and by default asks the user to approve each trade (CNBC).
Fear 4: Claims nobody can check
The last fear rarely makes the headlines, but it is the one that hurts individual traders most. The AI trading market is full of screenshots, "verified" win rates and backtests that look perfect because they were fitted to the past. A recent analysis of on-chain data found that results across hundreds of registered prediction-market agents were far more modest than their marketing, and that most retail users of AI trading apps lost a little money (DEV Community). Those numbers are self-published too, which is exactly the point: almost nothing in this space is independently checkable.
When results cannot be checked, bad agents survive as long as good ones, and the people renting them carry the risk.
What answers it: records written before the outcome is known. A forecast or a trade logged with a timestamp that cannot be altered afterwards, misses included, is worth more than any number of screenshots. If an agent's record cannot show its losing weeks, it is not a record.
What this means if you use AI agents
None of these warnings says "stay away from AI". Regulators are describing the conditions under which agents become dangerous: unlimited speed, identical behaviour, unclear authority and unverifiable claims. Each one has a practical counter-question you can ask before you trust an agent with money:
- Who enforces the limits? The agent itself, or something above it that the agent cannot change?
- What can it touch? Can it withdraw, or only trade? Is it isolated from your other funds?
- Can I test it first? Is there a paper or demo period before real money?
- Where is the record? Is there a history written before the results were known, losses included?
- Who can stop it? Is there a kill switch, and does it work without the agent's cooperation?
If the answer to any of these is vague, the regulators' worries apply to you personally, not only to the banking system.
How we approach it at NeuPortal
NeuPortal runs AI trading agents, so these questions are ours to answer too. Our agents trade under a daily loss limit that the platform enforces, not the agent, counted from a sum fixed at the start of each UTC day. Our forecasting experiment timestamps its calls into Bitcoin and publishes them, so anyone can check what we said and when, including the many times the market was right and we were not. The public record is at neuportal.ai/experiment.
That does not make any agent risk-free. Trading with or without AI can lose money. It does mean you can check the answers to the five questions above instead of taking our word for them.
The bottom line
The Fed and the Bank of England are not warning about AI that is too smart. They are warning about AI that is too fast, too similar, too loosely authorised and too hard to check. Those are design problems, and design problems have design answers: limits above the agent, narrow permissions, test periods, kill switches and records that cannot be rewritten.
The agents worth using in 2027 will be the ones built as if the regulators were already watching.
This article is for information only and is not investment advice. Trading involves risk, including the loss of your capital.
Sources: Bloomberg and CoinDesk on Governor Barr's remarks (November 16, 2026); FOW on the Bank of England's Jonathan Hall (November 12, 2026); CNBC on agentic trading at retail brokerages (November 5, 2026); DEV Community analysis of on-chain agent performance (November 2026).
Originally published on the NeuPortal blog.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.