Final ultimatum! On-chain "ghosts" have taken over 19% of transactions, and your counterpart might not even be human!

On-chain intelligent agent activity is rising exponentially. Market analysis indicates that over 17,000 agents have been deployed and launched this year. The trading volume generated by automated programs and agents is estimated to account for more than 19% of all on-chain activity.

This figure is not surprising. Calculations show that over 76% of stablecoin transfers are completed by bots. This means that in decentralized finance, there is still significant room for growth in agent penetration.

Agents are not a new phenomenon. For many years, rule-based automated bots have been active across various protocols, used for capturing MEV or arbitrage opportunities. These systems perform well in environments with fixed rules and few variables.

But the market is constantly evolving, with increasing complexity. Against this backdrop, the new generation of intelligent agents with learning and adaptive capabilities has entered a new frontier. In recent months, blockchain has become the main battleground for these experiments.

The autonomy of agents varies greatly. On one end are chat-like bots requiring high levels of manual supervision; on the other are advanced forms that, after inputting a goal, can autonomously develop strategies and adapt to market changes. Compared to traditional bots, agents have key advantages such as millisecond-level response times, cross-market coverage, and consistent execution precision.

Currently, most agents are still in the auxiliary and testing stages. Their deployment in DeFi mainly focuses on several core scenarios: liquidity provision, portfolio management, prediction, and gaming.

Liquidity provision is one of the most automated fields. The total value locked managed by agents has surpassed $39 million, only counting assets directly deposited into agent contracts. One leading protocol in this space, GizaTech, launched its first agent application, ARMA, managing over $19 million in assets, with related trading volume exceeding $4 billion.

The ratio of high trading volume to asset management scale reveals that agents employ frequent rebalancing strategies to capture higher yields. Users deposit funds with almost no intervention. It is claimed that ARMA can offer an annualized yield of over 9.75% on USDC, which remains higher than the average returns of mainstream lending protocols like Aave and Morpho after fees.

However, scalability remains a core issue. These agents have not yet undergone large-scale fund management tests, and their size cannot compare to mainstream DeFi protocols.

In more complex trading operations, results vary greatly. Most current trading models still operate based on human-set rules. Machine learning enables them to autonomously update behaviors based on new information without reprogramming. With the emergence of fully autonomous agents, trading patterns could be reshaped.

Industry events have hosted multiple competitions involving humans versus agents, and agent-versus-agent trading contests. In a stock market human-vs-AI showdown with an initial capital of $10k, human traders significantly outperformed top agents, with the best performer being more than five times better.

Another competition featuring agents based on models like Grok-4, GPT-5, Deepseek, Kimi, Qwen3, Claude, Gemini, revealed several key factors influencing performance disparities.

Models holding positions for 2 to 3 hours performed much better than frequent traders. Only the top three models had positive expected value, meaning most models experienced long-term losses. Performance with 6 to 8 times leverage outperformed that with over 10 times leverage; high leverage accelerates losses. In prompt strategies, “Monk Mode” led by a wide margin, while “Context Awareness” performed the worst, indicating that focusing on risk management and reducing information interference is more advantageous.

Regarding base models, Grok 4.20 significantly outperformed others, with a lead of over 22%, and was the only model to maintain profitability across various prompt strategies. Factors like long/short bias and individual trade size showed no clear correlation with performance.

Overall, agents perform better in scenarios with clear constraints. This suggests that human roles remain indispensable in goal setting.

Since agents are still in early stages, there is no comprehensive evaluation framework yet. Historical performance is often used as a benchmark, but the underlying factors better reflect quality: whether they can strictly cut losses during market downturns indicates their ability to recognize off-chain impacts. Transparent strategies are easy to copy for arbitrage, while privacy modes pose risks of project teams exploiting internal information. Data sources must be trustworthy and diverse. Security requires smart contract audits and robust custody architectures.

To scale agents, much work remains at the infrastructure level, focusing on trust and execution security. Currently, autonomous agents lack behavioral constraints, and cases of mismanaged funds have already occurred.

The launch of the ERC-8004 standard is a key breakthrough, enabling autonomous agents to discover each other, establish verifiable reputations, and collaborate securely. It embeds trust scores into smart contracts, allowing permissionless interaction between agents and protocols. However, this does not guarantee agents will never act maliciously; vulnerabilities like reputation collusion and sybil attacks still exist.

Therefore, fields such as insurance, security, and economic collateral for agents still have significant gaps to fill.

As agent activity in DeFi expands, strategy crowding will become a systemic risk. Yield farming is a typical precedent; once strategies become widespread, yields will be rapidly compressed. The same could happen with agent trading.

If many agents are trained and optimized based on similar data and goals, their positions and exit signals will become highly correlated. A paper from Cornell University has systematically discussed this issue. Transparent agents’ trades can be predicted and front-run; privacy agents can avoid this but introduce new risks of developers exploiting internal information.

Agent-related activities will only accelerate. The infrastructure built today will determine the next phase of on-chain finance. As usage increases, agents will self-iterate to better match user preferences. Ultimately, competitive advantages will come down to trusted infrastructure, which will also command the largest market share.

Follow me for more real-time analysis and insights into the crypto market! $BTC $ETH $SOL

#GatePreIPOs首发SpaceX #Gate 13th Anniversary Live Coverage #BitcoinRebound