Recently, I’ve been pondering a question: Will AI replace traders? Actually, this question is already outdated. What truly deserves attention is that the financial markets are engaging with a whole new form of participation.

It’s not just automation, but autonomy. These intelligent agents can not only process information, generate strategies, and place orders, but also hold wallets, invoke payments, manage positions, pay operational costs, and continuously execute within authorized scopes. They are designed and funded by humans, but in trading interfaces, they are already beginning to exhibit characteristics of independent economic units.

This sounds very futuristic, but the infrastructure is already in place. In March this year, the U.S. SEC issued a new interpretation of digital assets, and Nasdaq was approved to advance security tokenized trading. At the same time, Visa expanded USDC settlement to institutional scenarios in the U.S., with an annualized scale exceeding $3.5 billion; Coinbase launched USDC-settled perpetual contracts for non-U.S. users. These may seem like small changes individually, but combined, “machines directly participating in financial markets” has moved from imagination into the preparation stage.

Why will the market inevitably accept these autonomous intelligent agents? Simply put, because of efficiency. 24/7 trading, seamless cross-timezone switching, real-time processing of massive information—these capabilities inherently hold commercial value for investors. Brokers, clearinghouses, custodians will also gain new business increments as a result. Once the market recognizes that autonomous participants can bring lower friction, higher turnover, and finer risk control granularity, evolution will continue regardless of anyone’s will.

But there is a hidden risk structure here. If autonomous agents rely on continuous revenue to pay for computing power, data, APIs, Gas, and custody costs, then under market pressure, the first to be affected won’t be their book value, but their sustainable operation capability. This could lead them to prefer high-liquidity, high-leverage positions, shifting their risk appetite.

Adding another layer, if many agents access similar information sources, model architectures, and execution templates, the market will be more prone to synchronized position building and unwinding than in the past. A single signal change causes all agents to “rationally respond,” but this concentrated rational response could evolve into rapid local liquidity disappearance. In a 24-hour, cross-platform, multi-collateral environment, such stampedes will be faster and more dispersed than in traditional markets.

There are also cross-layer contagion risks involving stablecoins, collateral, and RWA. Once stablecoins take on cash-layer functions, financial assets become collateral, derivatives margins are linked to on-chain lending, local risks can more easily propagate across layers. The recursive expansion brought by automated payments—automatic renewals, rebalancing, redeployments—if lacking upper limit controls, could evolve into hidden leverage.

The most fundamental issue remains the responsibility chain. When something goes wrong, developers say they only provide models, platforms say they only provide interfaces, custodians say they only execute instructions. This is not unfamiliar in traditional markets, but as execution speeds up, participants become more dispersed, and cross-jurisdictional trading increases, unclear responsibility shifts from legal issues to market structure problems. Institutions won’t entrust large funds to a system that fails without anyone accountable.

Looking ahead, three paths may emerge. The first is cautious expansion—regulators opening low-permission, low-leverage participation in permissioned environments—most likely forming a sustainable institutional foundation, but not the fastest growth. The second is a dual-track approach—compliance tracks and shadow markets coexist, with risk transmission but differing responsibility standards. Based on historical experience, this is probably the most realistic scenario in the coming years. The third is expansion preceding governance—markets grow first, with licensing, insurance, and auditing lagging behind, and only addressing issues after typical incidents occur.

Where should true governance start? Protocol layers need to address authorization standards, risk budget interfaces, payment policy engines, trade receipts, and anomaly halts. Institutional layers need to handle access, permissions, and onboarding. Regulators should focus on identity, responsibility, and continuous reporting. The key is enabling different markets to recognize the same responsibility chain, rather than operating in silos.

My judgment is that there will not be a market completely divorced from traditional finance, solely governed by AI. More likely, it will be a parallel subsystem composed of human capital delegation, machine execution, stablecoin cash layers, tokenized asset interfaces, programmatic clearing, and embedded regulation. It will initially operate in localized scenarios and then gradually expand.

The true watershed of this transformation lies in the increasing maturity of AI and Web3 infrastructure, as autonomous agents demonstrate powerful information analysis, financial algorithmic capabilities, and real-time risk control. As infrastructure improves, relative to traditional markets, they will become more flexible, with lower friction costs and richer cross-space investment portfolios. When capital, payments, authorization, clearing, and responsibility chains operate under the same set of rules, the主体性 of financial agents will grow stronger and stronger.

In this rapidly evolving AI era, this is not just a technological upgrade but a fundamental change in the structure of the global financial markets.