Anthropic Wants Ethics Built Into AI Agents That Bargain. AI Trading Bots Are Next in Line

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 Financial Times reported in May 2026 that Anthropic wants lawyers' negotiating ethics built into Claude agents that bargain against other agents (Crypto Briefing, May 2026). At a glance that reads like a legal-tech curiosity about contract bots. For anyone who funds an automated strategy, it is a trading story first. The closest thing a retail trader owns to an agent that bargains against other agents is an AI trading bot sitting in a live order book, and the newest of those bots no longer just fire a fixed rule. They adjust price, size, and timing in response to whatever the other side is doing. That places this article squarely in the AI trading bot sub-niche, and specifically in its agentic tier, the fastest-moving corner of the automated trading market in 2026. We have benchmarked our 2026 testing framework against adaptive engines, including Zephyr AI, for precisely this reason. The news forces a question our funded-account work has been circling since we began the program in 2020: when two bots face each other, whose rules of engagement apply?

What did Anthropic actually propose?

The FT framing is deliberately narrow. Anthropic is not publishing a trading rulebook. It is arguing that when one AI agent negotiates against another AI agent, the humans on both sides should agree on a shared baseline of conduct before the negotiation starts, and that baseline should live inside the agent rather than in a policy document nobody opens. Crypto Briefing reports the move could reshape industry standards and influence how AI governance develops more broadly. Prediction markets have already begun pricing Anthropic's own trajectory around this kind of positioning, with valuation contracts trading live on Polymarket.

Strip away the commercial framing and the core idea is plain. An autonomous agent optimises for an objective. If the objective is to win, the agent will do things that are simultaneously legal, effective, and corrosive unless someone constrains it. Proprietary desks learned that lesson during the flash crash era with dumb, rule-based algorithms that carried no malice and no brakes. Anthropic is trying to pre-empt the same lesson one layer up, at the model itself, before the agentic wave reaches retail trading accounts in force.

Why agent versus agent behavior is already a trading problem

Here is the part the coverage largely skipped. The practical danger in markets is rarely a single rogue bot doing something obviously illegal. It is two or more individually well-behaved bots producing a pattern in aggregate that a regulator later reads as manipulation. Two agents independently probing the same thin book can, with no coordination whatsoever, generate something that resembles layering or quote stuffing. Scale that across a dozen funds running similar models trained on similar data and you get correlated behaviour that no single participant intended and none of them can fully observe.

That reframes the Anthropic ethics debate for traders. The question is not intent, which is comparatively easy to police, but emergent pattern, which is hard to write into a rulebook and harder still to audit after the fact. When we log the decision trail from an adaptive strategy inside our live-trading evaluation framework, we can usually reconstruct what a single bot did and why, and we can usually pin the reasoning to specific order tags. Reconstructing why a cluster of independent bots converged on the same quote at the same millisecond is a different order of problem, and it is the problem agentic trading is quietly importing into retail accounts long before any regulator writes a rule for it.

What does an AI trading bot actually trade?

Strip the marketing language off most retail bots and you find four moving parts: a signal that decides direction, a filter that decides whether to act, a sizing rule that decides how much, and an exit rule that decides when to stop. Fixed-rule expert advisors on MetaTrader, signal relays, copy-trading products, and the newer agentic frameworks differ mainly in how those four parts get generated. A copy-trading product inherits them from a human source. A conventional bot hard-codes them. An agentic bot generates them, which is exactly where the Anthropic debate bites, because generated decisions are the decisions that most need an ethics layer wrapped around them.

Our testing framework documents those four parts before we fund anything, then compares the documented spec against the order log for the opening weeks of live trading. The aim is not to catch a bot cheating. It is to catch the ordinary drift that appears when a vendor ships an update that quietly changes how the sizing rule behaves, a mismatch we now treat as a red flag rather than a footnote.

How big is the gap between backtest and live?

We treat every backtest number as a hypothesis, never a fact. The gap between backtest and live is structural, not accidental. It comes from three places: costs the backtest understates, fills the backtest assumes but never proves, and regimes the backtest data simply never contained. A strategy backtested through a calm 2024-2025 sample can look robust right up to the first genuine liquidity event, at which point the assumptions that produced that smooth equity curve stop holding.

We will not repeat vendor performance figures as if they were our own. Specific win rates, profit factors, and drawdown percentages vary by strategy parameters and by the broker feed a bot is pointed at. The honest instruction is to verify them directly with the provider and to insist on a live, fee-inclusive track record rather than a modelled one. What we can say with confidence, after six-month live trials on roughly 50 platforms since 2020, is that the bots which survive our process are rarely the ones with the prettiest backtest. They are the ones whose live behaviour stays inside the band their own documentation predicted.

Where do drawdowns actually show up?

Drawdown is not a smooth line down. In the strategies we have run, the damage concentrates in a handful of event windows, and those windows are predictable in advance even when the magnitude is not. NFP, CPI prints, and FOMC decisions are the three recurring stress tests on our calendar. A bot with a sound sizing rule trades through them with a bounded loss. A bot whose sizing rule doubles down on a losing position, which is a common hidden behaviour, discovers its true risk exactly once.

This is where the Anthropic conversation connects most directly to a real portfolio. An agent optimising to win, with no constraint on how it sizes into a loss, will eventually size into a loss until the account cannot absorb it. The guardrail that matters to a retail trader is not a promise in a marketing deck, it is a hard limit the bot cannot talk itself out of. When we compare fixed-rule systems against adaptive ones, the adaptive systems tend to read regime better, and they also tend to respect a stop worse, unless that stop is enforced at the platform level rather than inside the strategy code.

What does the subscription model do to your edge?

Fees are where a strategy's edge gets quietly eaten, and the market leans on four common structures. A flat monthly subscription charges the same amount whether the bot trades or not. A performance or profit-share model takes a slice of net gains. A broker-spread-markup model funds the bot through a wider spread on every fill. A lifetime licence charges once, which shifts the vendor's incentive toward new sales rather than ongoing performance. Each structure interacts differently with a real account, and the interaction is rarely obvious until you model it against your own trade frequency.

Can you unplug a bot cleanly?

Disengagement is the least glamorous part of automation and one of the most important. A bot you cannot stop is a bot that owns your account. In our process we test the exit as deliberately as the entry: can we flatten open positions in a single action, does the vendor persist positions after a cancellation, and how long does a subscription take to actually terminate. The failure mode we watch for is a bot that keeps managing an existing position after you have asked it to stop, which turns a clean exit into a slow unsupervised one.

Connection handling matters just as much. What happens if the API connection drops mid-trade? A well-built bot has a defined behaviour for a lost connection: flatten everything, hold everything, or hand control back to the user, and it states which one before you fund it. A poorly built one improvises, so the behaviour depends on the day. We treat an undocumented drop policy as a disqualifying feature, not a minor gap.

Is any of this regulated?

For a sector that handles real money, the AI trading bot market is lightly regulated, and the honest starting point is that "regulated" almost never means the bot itself is licensed. What is licensed is the entity that accepts your funds, or the broker whose rails the bot uses. So the check that matters is not the vendor's marketing badge but the register entry behind it. Below is where to look and what each register actually proves. Where a provider cannot point to a specific register entry, the correct instruction is to verify directly with the provider's primary regulator rather than take a badge at face value.

Free Download: Claude Negotiating-Agent Due-Diligence Checklist: Vetting Anthropic Bargaining Bots Before They Trade Against Rival Agents
A pre-deployment checklist that stress-tests Claude-based negotiating agents on ethics guardrails, counterparty-agent tactics, disclosure rules, and escalation triggers before you let them bargain or trade autonomously.
Get the Claude Agent Checklist

Not sure which AI trading bot fits your strategy? Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

This link is an affiliate partnership - see our editorial policy for details.

How Zephyr AI compares

The dimension that separates Zephyr AI's adaptive position-sizing from the fixed-rule systems we have evaluated is drawdown control through event windows. Where a fixed-rule system discovers its real risk in the first NFP or CPI print it meets, an engine that re-sizes against realised volatility can cut exposure before the event rather than after it. That is a structural difference, not a cosmetic one, and it is the reason we keep it as a benchmark in our review cycle.

The same contrast applies to disengagement. On the single feature that matters most to a retail account, a documented single-action flatten and a clear termination path, Zephyr AI's process is the standard we now measure other bots against. We would still verify fee terms and any live performance claims directly with the provider, because our own figures vary by regime and no reviewer should present them as fixed.

There is a broader point here that ties back to Anthropic. The firm is arguing that ethics should travel inside the agent. For traders, the equivalent argument is that risk limits should travel inside the bot, not sit in a settings menu a user forgets. A strategy whose worst case is defined by the platform rather than by the model's mood is worth more than a strategy with a prettier backtest and no floor. That is the lens we bring to every bot we fund, and it is the lens the next wave of agentic trading tools will be judged against.

Not sure which AI trading bot fits your strategy? Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

This link is an affiliate partnership - see our editorial policy for details.

Try Zephyr AI: Top-Rated AI Trading Algorithm for 2026

Try Zephyr AI: Top-Rated AI Trading Algorithm 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 AI trading bot work under US Pattern Day Trader rules?

Managerial rules apply to the account, not to the bot. If your equity sits below the US Pattern Day Trader threshold, a high-frequency strategy can trigger restrictions regardless of who or what places the trades. Confirm the current threshold and your broker's rule set before funding anything fast.

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

Usually, but the prop firm's rules govern, not the bot vendor's marketing. Many funded programs restrict automated trading, news-window trading, or both. Read the funded account terms first, because a violation typically voids the payout, not just the trade.

What happens if the API connection drops mid-trade?

It depends entirely on how the bot was built, and the behaviour should be documented before you fund it. A defined policy says flatten, hold, or hand control back to you. If the provider cannot state which one applies, treat that as a warning sign.

How reliable are bot backtests?

Treat them as hypotheses, never as results. Backtests systematically understate costs, assume fills they cannot prove, and lean on price regimes that may never repeat. Insist on a live, fee-inclusive track record and check whether the vendor's own documentation predicted the live behaviour.

How do I check whether an AI trading bot provider is regulated?

Check the register, not the badge. Bot vendors are rarely licensed themselves, so the entity that holds your money is what matters. Use the FCA Register, ASIC Connect, CySEC, NFA BASIC, SEC EDGAR, or the MAS directory, and if no entry exists, verify directly with the provider's primary regulator.

Do these bots work on crypto as well as forex and futures?

The strategy logic often ports, but the execution environment does not. Crypto venues run 24 hours, use different fee schedules, and carry thinner books at certain hours. Performance figures from a forex test should not be assumed to hold on a crypto venue with different microstructure.

How much should a good AI trading bot cost?

Pricing varies widely by structure and by account size, so any single figure would be misleading. Compare the fee model against your trade frequency rather than against a headline monthly number, because a spread markup on a high-frequency strategy can cost far more than a flat subscription.

Can two bots running the same strategy affect each other?

Yes, and this is the underrated risk in the agentic era. Independently optimised agents probing the same thin book can create correlated behaviour that no single bot intended. That emergent pattern is precisely the kind of problem Anthropic is trying to get ahead of at the model layer.

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).

More in this category: AI Trading Bot Reviews.