Author: Frank, PANews

Overnight, it seems everyone is deploying crayfish (a popular cryptocurrency platform). This trend has finally reached the crypto industry. On March 3rd, Binance and OKX, two crypto giants, simultaneously launched and open-sourced AI Skills libraries for AI Agents, enabling AI Agents to directly achieve on-chain alpha discovery and real-time trading through these protocols. Shortly before, prediction market leader Polymarket also launched a CLI tool specifically for agents.

Behind this seemingly coincidental situation lies the fact that AI is becoming the main trading entity in the future of the crypto industry, and this change has already begun.

But the core question facing users is: are agent-based transactions really reliable?

Leading the way, crypto officially welcomes AI traders.

Let's take a look at what Skill, which Binance and OKX have open-sourced this time, can actually do.

Binance's seven Skills are positioned as a "unified intelligent core," transforming fragmented crypto market signals into actionable trading decisions. Specifically, they enable AI agents to automate spot trading execution, such as accessing real-time market data and placing orders. They can also analyze any wallet address to generate smart money tracking reports, including detailed holdings. Other features include token retrieval, copy trading, and contract risk monitoring.

OKX's OnchainOS AI upgrade is positioned as an "on-chain operating system for AI agents." It supports over 60 on-chain functions related to autonomous wallet management, transactions, and payments. These include features such as wallet holdings lookup (cross-chain asset balances and portfolios), DEX market data, trade execution, and token discovery.

Polymarket's Rust CLI interface, launched earlier, is a terminal for AI agents, allowing them to directly query, trade, and manage all prediction markets on Polymarket. Furthermore, Bitget and Coinbase have also released similar skill libraries.

From a functional perspective, these skills provide the basic functions that ordinary users need for on-chain transactions or participation in other crypto transactions, including market research, order execution, smart money tracking, and more.

However, does this mean that everyone can now enjoy coffee while watching crayfish work behind the scenes to make money for them?

A user on social media shared a "crayfish" money-making tool.

AI Agent ≠ Quantitative Robot

But the actual result may be different from what most people imagine.

Many people equate "AI trading" with quantitative trading robots, but the underlying logic of the two is fundamentally different.

The difference is fundamental. Traditional quantitative trading robots are essentially automated programs that execute pre-defined rules, such as "buy when RSI falls below 30 and sell when it rises above 70." They are extremely fast, but they have no understanding of what they are doing, cannot read the news, and are unaware of market sentiment. The effectiveness of their strategy depends entirely on the person who wrote the code.

At the heart of the AI Agent is a large language model. It can read a news article about the Federal Reserve raising interest rates, understand what this means for the crypto market, and then decide whether to reduce its holdings.

Simply put: The bot executes the rules, and the agent makes the judgments.

In other words, the current agent doesn't monitor the market itself and then place an order directly when an opportunity arises. The resulting token costs and time lag are devastating for trading.

Current agent trading tends to adopt a "division of labor" model: traditional programs are responsible for monitoring and execution, while large models are only responsible for analysis and decision-making.

Specifically, a traditional program continuously pulls real-time prices, on-chain data, news, and other information from the exchange, then packages this data and sends it to a large model. The large model integrates multi-dimensional information such as market conditions, news, and on-chain anomalies to provide a trading decision, such as "Buy ETH, 10% position, order price $2450". Finally, the trading instruction is returned to the traditional program, which executes the order through the exchange interface and continuously tracks the results.

Traditional code acts as the agent's "hands" and "eyes," while the overall model serves as the "brain." The Skills offered by the three major platforms essentially provide the agent with standardized "hands" and "eyes," allowing it to quickly access the data and trading capabilities of various trading platforms. However, behind the scenes, humans still design the trading logic based on specific strategies. It's not about simply connecting to the Skill and watching your account balance automatically skyrocket.

Beyond technology and functionality, there are two real-world issues that must be addressed.

The first is speed. Traditional high-frequency quantitative bots have trading latency in the microsecond to millisecond range, with professional systems even achieving sub-millisecond latency. The key bottleneck for AI agents, however, lies in the time required for large-scale model inference. A complete analysis and decision output typically takes between several hundred milliseconds and several seconds, and in complex scenarios, it can even exceed 5 seconds. This is thousands or even millions of times slower than traditional bots.

Therefore , agents simply cannot compete with quantitative bots in terms of speed . They cannot perform high-frequency arbitrage or profit from millisecond-level price differences. The competitiveness of agents lies in the quality of their decisions : a quantitative bot can place an order in milliseconds, but it doesn't know the meaning of "the Federal Reserve Chairman just sent a dovish tweet," while an agent does. Agents are better suited to making one or two well-thought-out trades per hour, rather than performing thousands of mechanical operations per second.

The second factor is cost. Traditional bots, once developed, only require server costs to run. However, agents call large model interfaces every time they make a decision, which incurs expenses. For example, with GPT-5.2, if an agent analyzes the market every 5 minutes (288 times a day), the monthly inference cost is approximately $106. Using the more powerful Claude Opus 4.6, it's around $238. This isn't a significant amount for traders managing large sums, but for retail investors with only a few thousand dollars in capital, this inference cost, combined with transaction fees, makes achieving a net profit much more difficult.

When it comes to making money as an agent, there are more pitfalls than opportunities.

In addition, the quality of the agent's decision-making is also a major issue. Behind those seemingly logical and clear judgments, there may very well be absurd decisions.

In 2025, an AI trading competition held by Nof1 provided a stark example. Multiple large-model-driven agents competed, with wildly divergent results: the GPT-5-driven agent lost 62% of its initial capital, while Qwen3 and DeepSeek achieved profits of 22.3% and 4.89%, respectively. In this AI trading competition, while some models ultimately profited, they exhibited extremely unstable characteristics. DeepSeek, in particular, demonstrated high returns initially followed by a significant drawdown, which dampened market expectations.

In the second season of the experiment, 15 AI bots, each with a principal of $10,000, participated. Only GROK-4.2 achieved a positive return. Overall, only three models achieved positive returns in both seasons, while the rest were in a loss-making state.

Furthermore, PANews also conducted simulation studies on several of the most powerful models at the time, and the final results showed that, in the long run, their expected profits were all negative. (Related reading: Quantitative AI Assessment: Expected Profits for All Models Less Than 1, How Far Is Artificial Intelligence from Replacing Traders? )

On Polymarket, the most classic AI bot strategy is mathematical parity arbitrage: when the total cost of buying both "yes" and "no" contracts in a binary market is less than $1, buying both simultaneously locks in risk-free profits. Many bloggers have highly praised this strategy. However, Polymarket has responded by introducing dynamic fees and other rule adjustments, rendering some arbitrage strategies ineffective.

Overall, agent trading is not a "money printing machine." Model selection, strategy design, and risk control discipline are all indispensable.

In addition to these, agent transactions also involve several other risks that need to be considered.

Firstly, regarding security, the agent holds the private key and executes transactions autonomously. If the operating environment is compromised, it could lead to asset loss. Previous cases have shown that malicious techniques have been injected into open-source platforms to steal user keys. All three platforms used cautious disclaimers in their statements, with Polymarket even directly labeling it as "early experimental software."

Secondly, the "illusion" problem of large models cannot be ignored. Models sometimes generate analyses that seem reasonable but are actually wrong. In everyday conversations, this may only be embarrassing, but in trading, it could mean a loss of real money.

The homogenization of strategies is also a cause for concern. When a large number of agents use the same skills and the same models to analyze the same market, their judgments become highly similar, buy signals are triggered simultaneously, prices are rapidly driven up, and the space for latecomers is squeezed out.

AI is merely a weapon; it is still human who wields it.

The rules of the game in the crypto market are undergoing a profound shift as exchanges begin designing products for agents rather than humans. Data from 2023 shows that automated systems already accounted for over 70% of trading volume in the crypto market, and this percentage is still rising.

However, agent trading is still in the "early experimental" stage. The underlying logic is that this is merely an improvement in the tool, not "automating profit generation." Don't forget that institutions with extensive strategy and quantitative experience are also using the same tools to make improvements.

For ordinary investors, instead of rushing to build their own AI agents, it's better to first restrain FOMO (fear of speculation) and understand their limitations and weaknesses. Admittedly, the era of agent trading has arrived, but profitability still depends on the strategic decision-making abilities of the humans behind it.

DeAgentAI Genesis Airdrop is available on OKX, with a prize pool worth $50,000

Popular Articles