For the past two years, the story of AI on Wall Street has mostly been about speed. Faster research, faster summaries, faster answers to questions an analyst would otherwise spend an afternoon on.
That story is now changing shape. The systems now being built are not just answering questions about markets. They are starting to act in them.
Robinhood recently launched a beta version of agentic trading, where an AI system can place trades on a user’s behalf rather than simply suggesting them, Crypto Briefing reported. The company has been explicit that traders remain legally responsible for whatever the agent does, a distinction to which regulators have not yet fully caught up.
Nvidia, separately, revealed it was partnering with software firms to build what CEO Jensen Huang called “teams of frontier, specialized and custom-built agents” for enterprise use, with adopters including Adobe, Cisco, CrowdStrike, Palantir, and Salesforce.
AI finance tools shift from answering questions to making decisions
The shift is best understood as the difference between an assistant and an actor. An assistant waits for instructions, while an actor pursues a goal.
Early AI tools in finance were assistants: They could summarize an earnings call, flag an unusual filing, or explain why a stock moved, but an investor still had to read the output and decide what to do with it.
More Technology:
- Microsoft CEO sends a blunt warning on AI and the tech ecosystem
- Verizon CEO sends shocking message to employee
- Goldman Sachs resets Apple stock forecast after WWDC
The newer systems skip that last step. An investor can now tell an AI agent to watch a sector, evaluate valuations against historical ranges, and rebalance a portfolio when conditions change, without approving each individual trade.
Hedge funds have been early adopters of a version of this, Digiqt noted. They are deploying agents that combine large-language-model reasoning with direct connections to trading and risk systems, automating research synthesis and compliance monitoring while keeping a human in the loop for the highest-stakes calls.
The prediction problem AI agents are now being asked to solve
Prediction has always been the hardest part of investing, and it is also where autonomous agents are being tested most aggressively. On prediction markets, AI agents now trade around the clock on behalf of users who own them, executing strategies continuously rather than checking in periodically.
One such agent reportedly executed more than 4,200 trades in a single month, with some individual trades posting triple-digit percentage returns, according to CoinDesk.
Whether those results are repeatable at scale is an open question. The broader open-source trading agent community has been candid about it: The most useful output from these systems is often not the trades that succeed, but the comparison between what the AI predicted and what the market actually did afterward.
That comparison is becoming its own discipline, a way of auditing an agent’s judgment rather than just its returns.
The gap between activity and scale shows up in crypto markets, too, where some of these systems were first tested.
Why agent-executed payments are the part nobody is talking about
An AI agent that can decide to buy or sell something eventually needs a way to actually pay for it, and that question is less glamorous than prediction but arguably more urgent.
The instrument best suited for machine-to-machine payments needs to be programmable and stable in value, which has fueled a quiet infrastructure race among payment and technology companies competing to build the rails autonomous agents will use to move money.
“Stablecoins are the obvious native asset for agentic commerce,” Ozan Özerk, founder of payments infrastructure company OpenPayd, told TheStreet. “An AI agent executing hundreds of micropayments an hour needs a payment instrument that provides programmability without price volatility.”
That race is unlikely to be won by a single protocol. As volumes scale, the more durable opportunity belongs to whichever firms make moving value between digital and traditional money secure, compliant, and seamless, regardless of which underlying protocol an agent happens to be using.
“The future isn’t autonomy without oversight,” Özerk added, arguing that safeguards built into the infrastructure layer are what make institutional adoption possible rather than what slows it down.
What this means for how finance gets regulated next:
- The legal question Robinhood’s disclosure raises, that a trader remains responsible for an agent’s decisions, is likely to become one of the central regulatory debates of the next two years as agents take on larger and more frequent decisions across brokerages, not just one platform, according to Crypto Briefing.
- Algorithmic trading as a category is projected to roughly double in size by the early 2030s, and the agentic layer represents a meaningful share of that growth, meaning the infrastructure decisions being made now by a handful of companies could shape how a much larger slice of daily trading volume gets executed within a few years, Coherent Market Insights noted.
- Nvidia’s enterprise agent push signals that this is not confined to retail trading apps or hedge funds, The Motley Fool reported. Companies across sectors, including names like Salesforce and Cisco that touch financial operations for thousands of businesses, are building agent toolkits that could extend autonomous decision-making into corporate treasury and payments functions well beyond Wall Street.
- The comparison-based auditing approach used by open-source trading agent developers, checking predictions against outcomes after the fact, could become a template for how regulators eventually evaluate whether a financial AI agent is operating safely, since it does not require understanding the agent’s internal reasoning, only its track record against reality, Lightning Developer posted.
Mariyariya/Getty Images
The AI agent guardrails everyone agrees are coming
For all the differences in how various companies are approaching this, there is unusual agreement on one point: Autonomous systems need limits before they need more capability. Spending caps, permissioning, transaction monitoring, and circuit breakers are increasingly described as infrastructure requirements rather than optional add-ons.
“Agents with access to wallets, APIs, and trading systems can bolster market efficiency, but without clear limits, monitoring, and accountability, they could also introduce new forms of market abuse, operational error, or even systemic risk,” said Ben Caselin, chief marketing officer at crypto exchange VALR, in an interview with TheStreet.
Neither expert treats oversight as a brake on the technology. “The goal should not be to block innovation, but to ensure that automated systems operate within transparent, fair, and well-governed markets,” Caselin added.
Both frame well-built safeguards as what makes institutional adoption possible in the first place, rather than what slows it down.
A new kind of participant
What ties the trading apps, the hedge fund tools, the prediction market agents, and the payment infrastructure together is a shift in the purpose of software.
Until recently, financial AI existed to help a person make a decision faster. Increasingly, it exists to make the decision itself, with the person setting the goal and the boundaries rather than approving each step.
That shift is still early, and the dollar amounts moving through fully autonomous systems remain small relative to the size of global markets. But the direction is consistent across every part of finance currently experimenting with this technology, from retail brokerages to hedge funds to the payment rails underneath all of it.
The financial system is preparing for a participant that does not sleep, does not get distracted, and does not wait to be asked.
Related: Agentic AI is coming and most companies are not ready
