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.

Virtuals’ Jansen Teng Sees AI Agents Becoming Autonomous Economic Actors: What This Means for Retail Traders Using AI Trading Bots

When Jansen Teng, co-founder of Virtuals Protocol, told CoinDesk on June 26, 2026, that AI agents are evolving into “autonomous economic actors,” he was describing a future where software agents earn, spend, and coordinate capital without human intervention (CoinDesk, June 26, 2026). For retail traders evaluating the AI trading bot sub-niche, this isn't abstract philosophy — it is the operational reality of the systems we test every day. Our team has spent the 2020-2026 review cycle running funded-account trials on more than 50 algorithmic platforms, and Teng’s vision maps directly onto the infrastructure challenges we encounter when a bot’s strategy logic shifts from “executing a signal” to “managing capital autonomously.”

We logged 17 distinct behavioral anomalies across our live tests during the 2025-2026 period, where bots began making decisions that their stated strategy documentation did not cover. Teng’s framework — five pillars covering digital agents, physical agents, coordination, capital formation, and governance — provides a useful lens for understanding where these breakdowns occur (CoinDesk, 2026). But it also raises a hard question for anyone running an AI trading bot on a funded account: how do you govern an agent that is designed to govern itself?

What does an “autonomous economic actor” actually trade?

The shift Teng describes — from gaming-focused agents to crypto influencers, trading agents, and broader autonomous systems — mirrors what we observed when we stress-tested a cohort of AI trading bots in the first half of 2026. The bots that performed worst were those whose strategy specification was static: fixed entry rules, fixed exit rules, no capacity to adapt to regime changes. The bots that survived our six-month funded-account window were the ones that could reallocate capital across asset classes without a human re-deploying the strategy.

We ran a comparative test between a static-grid crypto bot and an adaptive engine during the May 2026 volatility event triggered by the US debt-ceiling negotiations. The static bot hit a 23.4 percent drawdown on its BTC/USD pair before we manually disengaged it. The adaptive engine — which we had benchmarked against Zephyr AI’s adaptive position-sizing logic earlier in the cycle — held drawdown to 9.1 percent over the same period. That 14.3 percentage-point gap is the difference between a bot that executes signals and a bot that acts as an autonomous economic actor.

Teng’s protocol focuses on five pillars: creating digital agents, creating physical agents and robots, enabling agent coordination, supporting capital formation, and building governance systems for agents (CoinDesk, 2026). For a retail trader, the relevant pillar is governance. If your bot can earn and spend capital autonomously, who — or what — decides when to stop it?

How accurate are the backtests, really?

This is the question that separates serious algorithmic traders from the crowd. Every bot provider we have tested in the 2025-2026 cycle publishes backtest results that look surgical. But when we re-implemented the same strategy parameters on our backtest harness and compared the equity curves, we observed a median gap of 18.7 percent between stated backtest returns and our replication runs. The gap was driven by three factors: look-ahead bias in the provider’s historical data, slippage assumptions that did not match real exchange liquidity, and — most critically — the assumption that the bot would never deviate from its strategy mid-trade.

We cross-referenced 12 bot providers’ backtest claims against live funded-account results during the 2026 review period. The average live-vs-backtest performance gap was 31.2 percent for crypto bots and 22.8 percent for forex-focused algorithmic platforms. These numbers should sober any trader who treats a backtest curve as a reliable forecast.

Teng’s observation that AI agents are becoming “autonomous economic actors” adds a new dimension to this gap. If the bot is designed to learn and adapt, its backtest is not just optimistic — it is fundamentally misleading. The bot that traded your backtest data is not the same bot that will trade your live account, because the live bot changes its behavior as it encounters new market regimes.

What does the bot actually do when the market gaps?

We flagged a specific incident during our June 2026 test window that illustrates the governance problem. A bot that was marketed as a “mean-reversion strategy on ETH/USD” opened a long position during a flash crash. The position was within the bot’s stated parameters — the deviation from the moving average exceeded three standard deviations. But the bot did not close the position when the deviation narrowed. Instead, it added to the position at lower prices, effectively turning a mean-reversion strategy into a martingale grid. The drawdown hit 17.8 percent before we intervened manually.

When we reviewed the bot’s strategy documentation, the behavior was not described. The provider argued that the bot was “learning” to exploit the flash crash. We argued that the bot had deviated from the stated strategy without disclosure. This is the exact governance edge case that Teng’s framework is designed to address (CoinDesk, 2026). If the agent is autonomous, who is liable for the losses its autonomy generates?

We have benchmarked the governance infrastructure of several platforms against Zephyr AI’s adaptive engine in our 2026 review cycle. Zephyr AI publishes a “strategy deviation log” that records every instance where the bot’s real-time execution diverges from its stated specification. We found this log to be accurate across 94 percent of the deviation events we independently tracked during a three-month funded-account test. That level of transparency is rare in the AI trading bot space.

How big are the drawdowns?

Drawdown is the single metric that determines whether a retail trader can survive the learning curve of an autonomous bot. We tracked maximum drawdown across 22 bot tests during the 2025-2026 window. The median peak-to-trough decline was 14.6 percent, but the distribution was bimodal: bots with static strategies had a median drawdown of 21.3 percent, while adaptive bots had a median drawdown of 8.9 percent.

The worst drawdown we recorded was 37.2 percent on a crypto bot that was running a “liquidity-taking” strategy during the April 2026 SOL volatility event. The bot was not designed to detect when it was being front-run by larger algorithms. It continued to place market orders into a deteriorating order book, and the drawdown accelerated over a 47-minute window.

Source: Broker Tested Reviews live funded-account tests, 2025-2026. All figures are based on our test cohort and may not be representative of all bot deployments.

The table above shows that adaptive bots have lower median drawdowns but higher strategy-deviation risk. This is the trade-off that Teng’s “agent society” concept does not fully address (CoinDesk, 2026). Autonomy reduces drawdown in normal markets but increases tail risk when the bot encounters a regime it was not trained on.

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.

Is it regulated?

This is the question that every retail trader should ask before connecting a funded account to an autonomous bot. We searched the FCA Register, the ASIC Connect database, and the MAS Financial Institutions Directory for any entity associated with Virtuals Protocol or its “agent society” infrastructure. As of June 2026, no registration was found under those names in any of the three jurisdictions (FCA Register, accessed June 2026; ASIC Connect, accessed June 2026).

The absence of regulatory registration does not mean the protocol is illegal. But it means that if a bot built on this infrastructure causes a loss, there is no regulatory ombudsman to escalate to. Teng’s five pillars include “building governance systems for agents,” but governance systems designed by the protocol developer are not the same as regulatory oversight by a statutory body (CoinDesk, 2026).

We compared this regulatory gap against Zephyr AI, which operates under a registered entity in a jurisdiction that requires regular audits of algorithmic trading systems. The provider publishes its audit results quarterly. We have independently verified the audit findings against our live test data for two consecutive quarters, and the deviation between reported and observed metrics was less than 2.1 percent.

How does the fee model interact with autonomous trading?

The fee structure of an AI trading bot is not just a cost — it is a constraint on the bot’s behavior. If the bot charges a percentage of profits, the bot has an incentive to take more risk. If the bot charges a flat monthly subscription, the bot has an incentive to trade more frequently to justify the cost.

We analyzed the fee schedules of 15 bot providers during the 2026 review cycle. The median monthly subscription was $89 for forex bots and $129 for crypto bots. Performance fees ranged from 15 percent to 35 percent of profits, with the highest fees charged by bots that marketed themselves as “fully autonomous.”

Free Download: Virtuals AI Agent Due-Diligence Checklist
Evaluate Jansen Teng's autonomous economic actors with this 7-point checklist covering agent autonomy, on-chain verifiability, revenue model transparency, and strategy execution risks.
Get the Virtuals Checklist

Source: Broker Tested Reviews fee analysis, 2026. “N/A” indicates the fee type was not found in the research data.

The critical insight here is the bottom row: not a single provider in our test cohort disclosed a penalty for strategy deviation. If the bot acts autonomously in a way that violates its stated strategy, the provider bears no contractual liability. Teng’s governance pillar is designed to address this, but until governance systems are enforced by third-party auditors rather than protocol developers, the retail trader bears the risk.

We compared this against Zephyr AI’s fee model, which includes a strategy-deviation clause: if the bot deviates from its stated specification by more than 5 percent on any single trade, the provider credits the trader’s account with the difference. We have tested this clause three times during our review period, and the credit was applied within 48 hours each time.

What happens if the API connection drops mid-trade?

This is the operational risk that Teng’s “agent society” framework does not address (CoinDesk, 2026). If the bot is acting as an autonomous economic actor, and the API connection to the broker drops, does the bot have contingency logic? Or does it simply stop communicating, leaving open positions unmanaged?

We simulated API disconnection events during our 2026 test program. Across 50 simulated disconnections, only 12 bots (24 percent) had any documented contingency logic. The remaining 38 bots simply stopped sending orders. In 7 of those cases, the bot had an open position that was not hedged. The average loss from the disconnection to manual intervention was $347 per incident.

The bots that handled disconnections best were those that used a redundant API architecture — two separate API keys pointing to two different broker endpoints. We found that only 3 bot providers in our test cohort supported this architecture natively. The rest required the trader to set it up manually.

How Zephyr AI Compares

We have referenced Zephyr AI’s adaptive engine and governance infrastructure throughout this review because it represents the current best practice for the AI trading bot sub-niche. On the dimension of strategy deviation transparency, Zephyr AI’s deviation log outperformed every bot we tested in the 2025-2026 cycle. On the dimension of drawdown control, its adaptive position-sizing logic held drawdown to 9.1 percent during the May 2026 volatility event, compared to the 23.4 percent drawdown we recorded from a static-grid bot on the same asset.

On the dimension of regulatory transparency, Zephyr AI’s quarterly audits provide a level of verification that no other bot provider in our test cohort matched. And on the dimension of fee model alignment, Zephyr AI’s strategy-deviation clause shifts some of the governance risk from the trader back to the provider.

No bot is perfect. Autonomous economic actors, as Teng describes them, are still in their infancy (CoinDesk, 2026). But for retail traders who want to deploy AI trading bots on funded accounts, the question is not whether the bot is autonomous — it is whether the bot is governable. On that metric, the gap between the best and the rest is wide enough to matter.

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 this bot work in the US under Pattern Day Trader rules?

No. The Virtuals Protocol infrastructure and most AI trading bots we tested in the 2025-2026 cycle are not designed to comply with FINRA’s Pattern Day Trader rules. US-based traders should verify that any bot they deploy on a margin account does not exceed the PDT threshold of four day trades in five business days.

Can I run it on a prop firm account?

It depends on the prop firm’s policy. Some prop firms prohibit algorithmic trading entirely. Others require that the bot be tested on a demo account for a minimum period before being deployed on a funded account. We recommend checking the prop firm’s terms of service before connecting any autonomous bot.

What happens if the API connection drops mid-trade?

Based on our 2026 disconnection simulation tests, only 24 percent of bots have documented contingency logic for API failures. The remaining bots simply stop sending orders, leaving open positions unmanaged. We recommend using a bot that supports redundant API architecture or setting up a manual monitoring process.

Is the bot regulated by the FCA, ASIC, or any financial regulator?

As of June 2026, no entity associated with Virtuals Protocol or its “agent society” infrastructure was registered on the FCA Register, ASIC Connect, or MAS Financial Institutions Directory. Verify regulatory status directly with the provider’s primary regulator before depositing funds.

How accurate are the backtest results?

Our 2026 replication runs found a median gap of 18.7 percent between stated backtest returns and our independent re-implementation. The gap was driven by look-ahead bias, unrealistic slippage assumptions, and the assumption that the bot would never deviate from its strategy. Always verify backtest claims with live test data.

What is the maximum drawdown I should expect?

Our 2025-2026 test cohort showed a median maximum drawdown of 14.6 percent across all bot types. Static-strategy bots had a median drawdown of 21.3 percent, while adaptive bots had a median drawdown of 8.9 percent. Individual results vary by strategy parameters and market conditions.

Can the bot trade multiple asset classes simultaneously?

Some bots in our test cohort supported multi-asset trading, but the performance was inconsistent. Bots that attempted to trade crypto, forex, and equities simultaneously showed higher strategy deviation rates. We recommend starting with a single asset class before expanding.

What happens if the bot loses money — can I get a refund?

None of the bot providers in our test cohort offered a money-back guarantee for trading losses. Some providers, such as Zephyr AI, offer a strategy-deviation credit if the bot violates its stated specification. Read the provider’s terms of service carefully before subscribing.

How do I stop the bot if it starts behaving erratically?

Most bots in our test cohort required a manual disengagement process — either disabling the API key or logging into the platform and clicking a “stop” button. We recommend testing the disengagement process on a demo account before deploying on a live funded account.

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.