How Much Freedom Should AI Trading Agents Get? What OpenAI, Anthropic, Two Sigma and Bloomberg Are Really Debating

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.

The most useful thing to come out of the current industry conversation about AI agent autonomy is not the disagreement. It is the agreement. OpenAI, Anthropic, Two Sigma and Bloomberg are all circling the same conclusion from different directions, and the reporting from Crypto Briefing captures it: an agent with real permissions needs a control layer that lives outside the agent itself. For anyone running an AI trading bot or an algorithmic trading platform, that is not an abstract governance point. It is the entire ballgame.

We have run six-month live trials of AI trading bots and algorithmic platforms on funded accounts since 2020, across more than 50 systems. In that program the agent always has permissions. It can size a position, move a stop, cancel an order, and roll exposure without asking anyone. The question we keep returning to is whether the thing holding the leash is independent of the thing wearing it. In our 2026 review cycle we benchmarked against the Ellington AI trading platform specifically because it separates the strategy layer from the portfolio risk layer, and that separation is the only reason we could score it on drawdown behavior at all. Most bots we evaluate cannot be scored that way, because the risk controls and the signal generator share the same code path.

What the debate actually covers

Strip the vocabulary down and the autonomy discussion reduces to three questions that any retail trader should be asking about their own automation:

1. What can the agent do without asking?
2. Who can stop it, and how quickly?
3. What happens to open positions when the answer to question two is "nobody, right now"?

The four organizations in the source material approach those questions from very different seats. That difference matters, because the seat determines which failure mode each one is most worried about.

The source does not publish numeric limits, thresholds, or timelines for any of these positions, so we are not going to attach numbers to them. What we can say is that the practitioner seat, the one occupied by a firm that actually runs money through automated systems, tends to be the most conservative about granting discretion. That is not an accident. It reflects what happens when your own capital is on the other side of the decision.

What does an autonomous trading agent actually do?

Plain English matters here. A rule-based expert advisor on MetaTrader 4 or MetaTrader 5 reads a fixed condition and fires a fixed order. An autonomous agent is different. It interprets a goal, chooses among candidate actions, and can chain them. In practice that means the agent decides position size, entry timing, and sometimes whether to trade at all.

That flexibility is where the strategy specification problem begins. When we evaluate an AI trading bot, we ask for the written spec: what the agent may trade, at what size, with what stop logic, and under what conditions it stands down. Where a vendor will not publish that document, we cannot score strategy adherence, and we say so in the review rather than guessing. Across our last review cycle we logged every order an agent placed over each six-month window and compared it against the written specification line by line. Where no specification existed, the comparison was impossible, and that itself became a finding.

Open-source frameworks set a useful contrast. NautilusTrader and Backtrader both require you to write the strategy explicitly, which means the spec is the code and there is nothing to drift from. A hosted agent that markets "adaptive" behavior is the opposite: the spec is a marketing page and the behavior is whatever the model decides on the day. Neither approach is automatically better, but they fail differently, and a retail account experiences those failures very differently. An open-source framework fails loudly and locally. A hosted agent can fail quietly and remotely.

Why does autonomy change your drawdown math?

Because an agent that can add to a losing position is a different risk object than one that cannot. A fixed expert advisor with a hard stop has a bounded worst case per trade. An agent with discretion over sizing does not, unless the platform enforces a portfolio-level cap that the agent cannot override.

In our testing program we look for three things on the risk side: a hard maximum position size that sits outside the strategy, a portfolio-level exposure ceiling, and a documented kill switch. Where those exist, we can describe drawdown behavior with some confidence. Where they do not, the honest answer is that the drawdown is undefined, and no amount of backtest history repairs that. This is the single most common gap we find in AI trading bots marketed to retail, and it is the reason we weight risk architecture more heavily than returns when we score a system.

The comparison matters here. A discretionary agent running on a general-purpose crypto bot platform, 3Commas being a familiar example with its DCA and grid families, inherits whatever risk controls that platform exposes. Those controls are real, but they are account-level rather than portfolio-level, and they generally do not model correlated exposure across strategies. A platform built around portfolio-level risk control, which is the design principle behind Ellington, handles the same volatility regime with a different failure mode. Under our 2026 algorithmic testing framework, the distinction tends to show up most clearly in how simultaneous positions in correlated instruments are treated. We are not going to hand you a drawdown percentage, because the number depends entirely on the strategy class, and any provider's published figures should be verified directly with that provider.

Where the backtest and the live account part ways

Always. The gap is structural, not a bug, and it is wider for agents than for rule-based bots.

Three reasons. First, agents adapt to recent data, so their backtest is partly a fit to the backtest window itself. Second, agents change behavior when execution costs change, and backtests usually assume a flat cost model that no live venue offers. Third, agents that route across venues interact with liquidity that a backtest cannot fully model, particularly around news events.

We treat any published backtest as a hypothesis, not evidence. For an AI signal provider or an AI trading bot, the only number we weight heavily is what happened in a funded account over a full market cycle, and even then we want the order-level log rather than the equity curve. When we ran comparable momentum and mean-reversion strategies through our 2026 algorithmic testing program, the divergence between the simulated path and the funded-account path was consistent enough that we now discount simulated results by default. That is not a criticism of any one vendor. It is a property of the problem.

Free Download: AI Agent Autonomy Due-Diligence Checklist: Setting Freedom Limits for Your Trading Bot (OpenAI, Anthropic, Two Sigma & Bloomberg Frameworks)
A trader's checklist for grading how much unsupervised decision freedom an AI trading agent should get, covering human-in-the-loop triggers, kill-switch and rollback rules, order-size and exposure caps, model/version transparency, audit logging, and who is liable when the agent acts alone.
Get the autonomy checklist

If a provider cannot fill in that last row, treat the whole table as unverified.

The fee model question nobody asks

Subscription pricing interacts with strategy economics in a way that most reviews ignore. A flat monthly fee is punishing for small accounts and irrelevant for large ones, which means the same bot can be rational for one trader and irrational for another without any change in performance.

Performance fees create a different problem. They align the vendor with gross returns but not with drawdown-adjusted returns, so a provider paid on profit share has an incentive to let the agent take more risk than a client would choose. That is a structural conflict, and it is the kind of thing that never shows up in a marketing page.

Broker and API integration sits underneath all of this. MetaApi and similar bridge layers, MetaTrader's own API surface, and direct exchange APIs all behave differently under latency and partial fills. A strategy that is profitable on one execution path can be unprofitable on another, and the per-trade difference, while small, compounds quickly over hundreds of trades. For a retail portfolio, that compounding is the difference between a strategy that pays for its subscription and one that does not. Ellington's fee transparency is one of the few places in this market where the cost of the automation is stated in the same place as the cost of the trades.

What happens when the API connection drops?

This is the failure mode retail traders underrate most. An agent that runs in the cloud and talks to your broker through an API behaves differently from one that runs locally and dies with your machine. Both are legitimate designs. The problem is that vendors rarely document what the agent does in the gap: does it flatten, does it hold, does it retry, and how many times before it gives up?

In our live-trading evaluation framework we test disconnection behavior explicitly, and it is one of the few dimensions where we can give a clean pass or fail rather than a hedged note. If the provider cannot tell you the failover behavior in writing, assume the worst case, which is an open position with no stop management until the connection returns. For a leveraged account, that worst case is not a minor inconvenience. It is the scenario that ends accounts.

The same logic applies to disengagement. Can you actually stop the bot cleanly? In our funded test accounts we measure how long it takes to flatten every position and revoke the agent's permissions, and we measure it in actions rather than minutes. If that takes more than one documented action, we flag it. A system you cannot exit under stress is not a system you control.

Is any of this regulated?

Partly, and the gap is the interesting part. The vendor that sells you an AI trading bot may be regulated, may be unregulated, or may be regulated in a jurisdiction that has nothing to do with where you live. The primary sources are free and worth ten minutes of your time before you fund anything.

If a vendor will not tell you which register to check, treat that as the answer.

The under-discussed edge case is the tool-versus-manager line. A vendor can call its product a tool, but if the agent decides what to buy, when, and how much, without a client instruction for each decision, some regulators will look at that and see discretionary portfolio management, which requires authorisation the vendor may not hold. That is not hypothetical. It is the kind of thing that gets discovered after a bad quarter, when clients start asking who was responsible for the decision. Two Sigma, whose advisory entities are searchable in the SEC's Investment Adviser Public Disclosure database, operates inside a framework that assumes this question has been answered. A retail bot vendor selling an "autonomous" agent from a jurisdiction with light oversight has usually not answered it at all. Verify any licensing claim directly with the provider's primary regulator rather than relying on a badge on a website.

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.

How much rope should an agent get with your capital?

The useful way to think about it is in tiers, and to match the control to the tier rather than to the marketing copy. Most retail blowups we have studied come from a bounded-tier expectation meeting a discretionary-tier reality.

Anything not documented in the provider's own materials should be marked "verify with provider" rather than assumed.

How Ellington Compares

On the dimension that matters most in an autonomy debate, the comparison is not close. Portfolio-level risk control that sits outside the strategy is rare in this market. Most of what we test, from 3Commas-style crypto bot platforms to single-purpose expert advisors, puts the risk logic inside the same process that generates signals, which means a bug or a regime shift can affect both at once. Where Ellington's multi-strategy automation separates the two, the agent can be aggressive inside a mandate without the mandate itself being negotiable. That is precisely the design question OpenAI, Anthropic, Two Sigma and Bloomberg are circling, and it is the one we score hardest. The second concrete edge is coverage: a single agent that spans multiple asset classes under one risk layer is easier to supervise than four separate bots with four separate kill switches, and supervision is the whole point of this article.

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.

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.

Frequently Asked Questions

Does an autonomous AI trading agent work in the US under Pattern Day Trader rules?

PDT rules apply to margin accounts and restrict accounts that place four or more day trades within five business days unless minimum equity requirements are met. An agent that trades frequently can trip that threshold quickly, so verify the current requirement with your broker before funding. See FINRA's day trading guidance for the primary rule text.

Can I run an AI trading agent on a prop firm account?

It depends entirely on the firm. Many prop firms restrict or prohibit fully automated execution, and some require you to disclose the strategy and the tools you use. Read the terms of service before paying a challenge fee, because a violation can void the account and the fee.

What happens if the API connection drops mid-trade?

It depends on where the agent runs and what failover the provider has documented. Ask for the written behavior, and if there is no answer, assume open positions continue without stop management until the connection returns. That assumption should shape your position sizing.

How do I know the agent is following its stated strategy?

Ask for the written specification and order-level logs, then compare them. Where a provider will not publish the spec, we decline to score strategy adherence, and you should treat that silence as a risk factor rather than a neutral omission.

Are AI trading bots regulated?

Sometimes, and not always in your jurisdiction. Check the FCA Register, ASIC Connect, SEC IAPD, NFA BASIC, the CySEC list or the MAS directory depending on where the vendor operates, and verify directly with the provider's primary regulator before relying on any licensing claim.

Do autonomous agents beat simple rule-based expert advisors?

There is no reliable public evidence that they do, and comparisons are usually unfair because the backtests run on different windows with different cost assumptions. Treat autonomy as a feature with costs attached, not as an edge in itself.

How much capital should I put into an agent?

Only what you can afford to lose, and start smaller than you think is sensible. Our testing program uses funded accounts precisely so that the evaluation does not depend on the vendor's own reported numbers.

Can I stop an agent cleanly?

Test the kill switch before you fund the account, not after. If the provider cannot show you how to flatten positions and revoke permissions in one documented action, that is a finding worth acting on.

What fees should I expect?

Usually a subscription plus execution costs, and sometimes a performance fee on top. Verify the schedule directly with the provider, and model the subscription against your expected trade frequency, because a flat monthly fee is punishing for small accounts and irrelevant for large ones.

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.

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: AI Trading Bot Reviews.