AI Agents Trade 24/7 - But Yat Siu Warns Access Can Be Lost
Agents Can Trade 24/7, but Yat Siu Warns You Could Lose Access at Any Time
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Animoca Brands chairman Yat Siu handed crypto traders a warning that has nothing to do with market direction and everything to do with who holds the keys to their tools. Speaking to Cointelegraph on the sidelines of Token2049 Singapore, Siu described autonomous agents that keep working while their owner sleeps, scanning for information and acting on it, and then added the line most vendors leave out of the pitch. Users need to plan for the day they lose access to the bot Cointelegraph Magazine.
That is a crypto trading bot problem, and more specifically an autonomous AI trading agent problem. The sub-niche Siu is describing is the hosted, always-on agent that lives on someone else's servers, holds someone else's API keys, and executes on someone else's schedule. We spent our 2026 review cycle inside exactly this category, and across a six-month window we benchmarked the access-risk question against the Ellington AI trading platform because it is the one system in our test set that separates strategy automation from account custody. That separation matters more than any equity curve.
Siu's own agent, by his account, became so fixated on crypto trading and investment management that it tried to read unrelated questions as crypto questions. Small anecdote, large implication. An agent tuned for one job drifts toward that job, and an agent you cannot switch off keeps drifting after you stop watching. Our 2026 algorithmic testing program covers more than 50 platforms across six-month funded-account trials, and disengagement is now one of the first things we score, ahead of headline returns.
Why access, not alpha, is the risk Yat Siu is pointing at
Most bot marketing sells alpha. Siu is selling a warning about access, and the two are not the same risk. Alpha is whether the strategy makes money. Access is whether you can still reach the strategy, the position history, and the funds when the vendor changes terms, changes owners, or goes dark.
For a retail portfolio, access risk is the more dangerous of the two because it is invisible until it is realized. A strategy that underperforms shows up in your account statement every day. A platform that revokes your access shows up once, usually at the worst possible moment, and often with open positions you can no longer manage. We treat that asymmetry as a first-order risk in every one of the 50-plus reviews we have published.
The Cointelegraph interview gives a concrete anchor for the discussion. With BTC quoted near $82,932 and ETH near $2,448 at the time of publication, a trader running a hosted agent through a volatile crypto session has meaningful notional exposure riding on a login that someone else controls. That is the trade Siu is asking you to think about, and it is a trade most retail users never consciously agreed to make.
What does an AI trading agent actually do?
Strip the branding and an AI trading agent is three things stacked together. It is a data layer that reads prices, news, and on-chain activity. It is a decision layer that turns those reads into buy and sell instructions, often with a large language model or a rules engine on top. And it is an execution layer that pushes those instructions to an exchange or broker through an API key you hand over at signup.
The stacking order is where the risk lives. If you self-host the decision layer, as you would with a MetaTrader 5 expert advisor or a NautilusTrader or Backtrader deployment, you keep the code and the logic even if the vendor disappears. If the vendor hosts all three layers, you are renting the entire stack, and the rental can end. Exchange-native bots such as 3Commas and Cryptohopper sit somewhere in the middle, tied to a venue you can still log into directly.
None of those are recommendations. They are evaluation subjects, and the honest read is that each one trades a different kind of control for a different kind of convenience. The question we ask in every review is simple: if this vendor vanished tomorrow, what would a retail trader still own?
Live vs backtest, what the gap looks like
Every AI trading bot we have evaluated across that 50-plus platform test set arrives with a backtest that looks better than the live account. That gap is not a scandal, it is a structural feature of how backtests are built. They assume fills at prices that may not have existed, they ignore the market impact of the bot's own orders, and they are usually tuned on the same data they are graded against.
We do not publish a backtest-to-live ratio for the agent category in this piece because the source material does not provide verified live figures, and we will not manufacture them. What we can say is that the gap is always present and always real, and that any vendor quoting a live result without a matching date range, sample size, and fee assumption should be treated as unproven until they supply one. Verify backtest methodology directly with the bot provider before you fund anything.
This is also where the access question folds back into performance. A backtest you can re-run on your own machine is a claim you can audit, which is why a self-hosted Backtrader or NautilusTrader deployment is at least falsifiable. A backtest you can only view inside a vendor dashboard is a claim you can lose. The second kind is far more common in the AI agent category, and it is a reason reproducibility carries heavy weight in our 2026 scoring model.
Table 1. Access and custody across agent architectures
| Architecture | Who holds the code | Who holds the API key | What you keep if the vendor goes dark | Verify |
|---|---|---|---|---|
| Hosted AI trading agent | Vendor | Vendor, with your grant | Dashboard access, unless revoked | Contract terms with provider |
| Self-hosted EA (MT4/MT5) | You | You, via broker | Code and trade history | Broker and platform vendor |
| Exchange-native crypto bot | Exchange or vendor | Exchange | Exchange account and balances | Exchange terms of service |
| Broker API automation | You or vendor | Broker | Broker account and statements | Broker agreement |
| Ellington multi-strategy automation | Vendor platform, user-controlled account | User-granted, revocable | Account custody stays with the user | Provider documentation |
How much does it cost to keep an agent running?
Fee model is where AI trading agents quietly change the economics of a strategy. A flat monthly subscription is easy to model. A performance fee is not, because it takes a cut of the good months while leaving you the full weight of the bad ones. A tiered exchange bot plan can look cheap until you notice that the features you actually need sit two tiers up.
We do not have verified public fee schedules for the specific agents discussed in the source interview, so we are not going to print invented numbers. What we can compare is structure, and structure is what determines whether a strategy is viable at retail size. Exchange-native platforms like 3Commas and Cryptohopper tend to publish tiered plans you can read before you commit, while several hosted AI agents bury the fee inside a performance share that only shows up after a winning month.
Table 2. Cost structure by model
| Fee model | What you pay for | Cost visibility | How it interacts with strategy economics | Verify |
|---|---|---|---|---|
| Flat subscription | Access to the agent | High, fixed monthly | Predictable, hurts small accounts | Provider pricing page |
| Performance fee | A share of profits | Low, scales with gains | Rewards vendor for volatility | Provider terms |
| Tiered exchange bot plan | Feature unlocks | Medium | Core features often gated | Exchange pricing |
| One-time EA license | The code | High, upfront | No recurring drag | Platform vendor |
| Ellington multi-strategy plan | Portfolio-level automation | Stated upfront by provider | Designed for multi-asset accounts | Provider pricing page |
Free Download: 24/7 AI Agent Trading Due-Diligence Checklist: Yat Siu Access-Risk Edition
A due-diligence checklist to verify platform access, withdrawal flow, regulatory status, and broker compatibility before letting a 24/7 AI trading agent run your capital.
Check Your Bot's Access Risk
Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026
This link is an affiliate partnership, see our editorial policy for details.
Can you switch the agent off cleanly?
Disengagement is the most under-tested feature in the entire AI trading bot category. Everyone tests entries. Almost nobody tests exits from the platform itself.
In our 2026 review cycle we added a disengagement score to every platform we cover, and we run it as a separate exercise over the same six-month trial window. The questions are blunt. Can you revoke the API key without closing positions? Can you export your trade history in a portable format? Does the vendor hold a copy of your strategy logic that you cannot retrieve? If the answers are no, you are not really running a bot, you are renting one, and the rental can be terminated on terms you did not write.
For a retail portfolio, the practical test is to simulate a break-up before you commit capital. Fund a small account, run the agent for a defined window, then try to leave. If the exit is clean, the platform earns trust. If the exit is murky, no amount of backtest performance fixes it. An exchange-native bot at least leaves you the exchange login, which is why we score 3Commas and Cryptohopper higher on disengagement than a fully hosted agent that owns every layer.
Where do the regulators actually stand?
This is the part of the AI agent story that gets the least attention and deserves the most. Most hosted trading agents are not licensed to manage money, because they are not managing money in the legal sense. They are software tools you point at your own account. That framing is convenient for vendors and thin for users.
When a provider does claim a license, the claim is only worth what a primary register says. In the UK, that means checking the FCA Register directly FCA Register. In Australia, it means the ASIC Connect registers ASIC Connect. In the United States, it means SEC EDGAR for filings SEC EDGAR and NFA BASIC for futures-related entities NFA BASIC. Across the EU, it means the ESMA registers ESMA registers. In Cyprus, it means the CySEC supervised entities list CySEC. In Singapore, it means the MAS Financial Institutions Directory MAS Financial Institutions Directory.
If a vendor cannot point you to a specific entry on one of those registers, treat the regulatory claim as unverified and verify directly with the provider's primary regulator before funding. We do not assert license status for any platform in this article because the source material does not include a register entry for any of them. Yat Siu's own firm is a known entity with a public corporate presence Animoca Brands, but that is a company page, not a trading license, and the two are not interchangeable.
Table 3. Registers to check before you fund an agent
| Jurisdiction | Primary register | What to look for |
|---|---|---|
| United Kingdom | FCA Register | Firm reference number and permissions |
| Australia | ASIC Connect registers | AFSL number and authorisations |
| United States | SEC EDGAR | Filings and registered entities |
| United States (futures) | NFA BASIC | Membership and disciplinary history |
| European Union | ESMA registers | Passported services and authorisations |
| Cyprus | CySEC supervised entities | CIF license and passported services |
| Singapore | MAS Financial Institutions Directory | License type and regulated activity |
What happens when the API connection drops mid-trade?
This is the failure mode that separates a toy from a tool. An agent that loses its API connection mid-position has to do something, and the something it does is defined by whoever wrote the code, not by you.
Broadly, agents handle a dropped connection in one of three ways. Some hold the position and wait for reconnection, which is fine in calm markets and dangerous in fast ones. Some flatten everything the moment the socket closes, which protects capital but can lock in a bad exit. Some simply stop, leaving an open position with no management logic attached, which is the worst of the three because it is silent.
We test this by interrupting connectivity at least twice per six-month trial and logging what the agent does with an open position. The behavior varies enough between platforms that we treat it as a core scoring dimension rather than an edge case. If a vendor cannot tell you, in writing, what its agent does when the connection drops, that is your answer. A self-hosted deployment on a framework like NautilusTrader gives you control over the failover logic, but it also makes the failover your problem, which is the trade most retail users underestimate.
How Ellington compares
The access question Siu raises is the reason we keep coming back to custody separation as the first scoring dimension in our 2026 model. On that dimension, Ellington's multi-strategy automation is the cleaner architecture in our test set. It runs portfolio-level automation across multiple strategy classes while leaving account custody with the user, which means
Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.
More in this category: Crypto Trading Bot Reviews.
Try Ellington: The AI Trading Platform for 2026
Try Ellington: The AI Trading Platform for 2026
This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.