Brokerages, startups, and individual investors are building AI agents that can oversee portfolios, execute trades, and manage financial tasks around the clock. The shift from concept to deployment accelerated in 2026, with CNBC reporting that major platforms now offer agent connectivity, startups are shipping dedicated investing assistants, and retail investors are testing general-purpose models like Claude as stock-picking tools.
“Effectively everybody has their own family office that is working 24/7 for them while they’re awake or sleeping,” Devin Ryan, head of financial technology research at Citizens, told CNBC. “This isn’t 10 years away. This is coming in the next few years.”
Ryan’s forecast comes with a specific number: he estimates agentic finance could increase transaction volumes by at least tenfold. A retail investor who currently trades roughly twice a month could trade 20 times a day under an agentic model. By end of 2027, he believes the majority of transaction activity by trade count on some platforms will come from agents.
The Product Landscape
Three different approaches to agentic trading are shipping simultaneously.
Robinhood opened its platform to third-party AI agents in May, allowing external AI assistants to connect with customer accounts and execute trades. The company introduced both Agentic Trading and an Agentic Credit Card, letting users instruct agents to rebalance portfolios, monitor themes like AI stocks, or execute trading strategies automatically. “Our mission has always been to democratize finance for all, and now, that mission extends to AI agents,” CEO Vlad Tenev said at the time.
Robinhood’s implementation includes structural guardrails. Agentic trading accounts are separated from main portfolios, limiting agent access to specifically allocated capital. The system sends notifications when trades occur and lets customers disconnect an agent immediately. Initial beta support covers stock trading, with options, cryptocurrency, and futures planned for later.
Brokerage firm Public is taking the opposite approach, building AI agents in-house rather than opening its platform to third parties. “What this era of agentic is doing … it goes away from just being able to research something by yourself and then make up your own ideas and then trade the way you’ve traded where it’s now becoming automated and where AI agents can actually execute investment strategies on your behalf,” co-founder and co-CEO Leif Abraham told CNBC.
Public requires users to review and approve an agent’s workflow before it carries out any investing tasks. “You still have the last word,” Abraham said. “The AI agent will not have its own mind. … It will only execute.”
Startup Podium Markets AI occupies a third position: an independent AI assistant called Ivy that analyzes a customer’s portfolio across multiple brokerage accounts and generates recommendations based on goals and risk tolerance, but stops short of autonomous execution. “The AI informs, but the human decides,” said co-founder and CEO Dirk Mueller-Ingrand. “The average investor still should be very much in charge of the final decision.”
Retail Investors as Beta Testers
While institutional platforms build structured products, retail investors have spent three years testing what general-purpose AI can do with market data.
The results are mixed. Obioha Okereke, a 29-year-old technology consultant in Georgia and founder of the financial literacy platform College Money Habits, built an agent using Anthropic’s Claude to search for undervalued stocks and options opportunities. “It was essentially just asking Claude to act as a hedge fund analyst to find undervalued stocks,” he told CNBC, adding that he still reviewed every recommendation before placing a trade. “I will always stand by AI being a tool as opposed to a replacement.”
Thomas Schlossmacher, a 31-year-old founder whose company Specialty Tokens builds AI systems for businesses, tested a trading agent after seeing claims online that AI could uncover profitable market patterns. He reported consistent losses. “I think if you’re using it for an automated system or relying on an agent to do it for you, you probably want a professional,” he told CNBC. “To blindly give an agent and say, ‘Hey, make me money,’ I think is kind of dumb.”
The divergence between Okereke’s cautious tool-use approach and Schlossmacher’s automated-loss experience illustrates the core tension in agentic finance: the gap between “assist me” and “act for me” is where most of the risk concentrates.
The Intent Translation Problem
Teaching an AI agent to buy or sell a stock is straightforward. Teaching it what an investor actually means is a harder problem.
An investor might instruct an agent to “grow my portfolio aggressively.” That instruction is ambiguous across multiple dimensions: does it mean accepting more volatility, concentrating holdings, using leverage through options, or tolerating a higher probability of loss? An agent can faithfully follow its interpretation of the instruction and produce an outcome the investor never intended.
This is not a theoretical concern. Citizens’ Ryan flagged it directly: “You have to make sure that the customer’s best interests are at the forefront. If the agent is not behaving as modeled or as you expect, that becomes a risk for the firm.”
The firms building agentic trading products are each drawing the control boundary in a different place. Robinhood separates agentic accounts from main portfolios, limiting exposure. Public requires explicit workflow approval before execution. Podium Markets AI stops short of execution entirely. The spectrum runs from full autonomy with guardrails to autonomous recommendation with human execution, and no industry consensus has emerged on where the line should sit.
Infrastructure as Prerequisite
Autonomous trading agents surface an infrastructure requirement that portfolio analysis tools do not: they need to run continuously, and downtime has financial consequences.
According to CNBC’s reporting, agents fail when hosting infrastructure goes down mid-task, whether from a laptop entering sleep mode or a network outage. This forces reliance on always-on infrastructure, a shift from the intermittent compute patterns of traditional retail investing tools.
The infrastructure challenge extends beyond uptime. The Motley Fool published analysis on July 28 examining how agent code execution is becoming an infrastructure problem distinct from model inference. As agents write and execute code autonomously, rather than waiting for LLM responses, platforms must create isolated computing environments (sandboxes), control access, preserve state between attempts, and shut down environments when tasks complete.
The analysis identifies key operational metrics for agent infrastructure: how many sandboxes one server can support, how quickly each task finishes, whether performance holds under full load, and how much memory each environment consumes. These metrics favor different processor designs and create new competitive dynamics among hardware vendors. Amazon has made this requirement explicit with AWS Lambda MicroVMs for isolated agent code execution, and Microsoft and Google are adding similar capabilities to their cloud platforms.
For trading agents specifically, the intersection of financial execution and infrastructure reliability creates a compounding risk: an agent that loses connectivity while managing an open position cannot respond to adverse price movement.
Enterprise Governance Diverges From Financial Services
Financial services firms are accelerating agent deployment at the same time that non-financial enterprises are blocking it.
Louis Lehot, a Silicon Valley deal lawyer at Foley & Lardner with 25 years of experience advising companies from startup to public markets, reported in Mondaq on July 27 that procurement departments at some of the largest US companies are refusing to let AI agents operate on their networks. The observation came from a CFO Executive Forum held at Foley’s Palo Alto office on July 22, where panelists confirmed that “AI adoption” claims in fundraising decks rarely survive due diligence scrutiny.
The divergence is instructive. Financial services firms have decades of experience with algorithmic and quantitative trading systems, regulatory frameworks for automated decision-making (Reg SHO, market access rules, best execution requirements), and infrastructure designed for continuous operation. They also have a direct, measurable upside from 24/7 autonomous operation that most enterprise software deployments lack.
Enterprise procurement teams, by contrast, face governance, audit, and liability concerns with fewer established precedents for autonomous agent operation. The gap suggests that agentic adoption will follow industry-specific trajectories rather than a uniform technology adoption curve.
The Liability Question
Stanford Law School’s CodeX center published a white paper in July establishing what it calls the “Phantom Agent” framework for civil liability when autonomous AI agents cause harm. The framework maps existing legal doctrines, including agency law, respondeat superior, and corporate attribution, onto the chain from developer to deployer to user.
The paper’s central argument is that courts do not need to determine whether an AI agent is conscious or possesses intent to hold someone liable for its actions. Existing legal infrastructure is sufficient.
For agentic trading, the liability question has specific contours. If Robinhood’s platform connects a third-party AI agent that executes a series of trades causing significant losses, the liability chain includes the model developer (Anthropic, OpenAI), the platform (Robinhood), the agent developer (potentially a third party), and the user who authorized the agent. Stanford’s framework suggests that existing doctrines can distribute responsibility across this chain, but no court has tested the framework against an autonomous trading loss.
The regulatory dimension adds another layer. FINRA and the SEC have extensive rules governing automated trading for institutions, but the application of those rules to retail investors using autonomous agents through consumer brokerage platforms is untested. The question of whether a retail customer who authorizes an AI agent to trade constitutes “algorithmic trading” under existing regulations has no settled answer.
The Speed of Product Versus the Speed of Law
Three facts define the current moment in agentic finance.
First, the products are shipping. Robinhood, Public, and Podium Markets AI represent three different points on the autonomy spectrum, and all three are live or in active beta. The infrastructure layer is evolving simultaneously, with cloud providers building dedicated agent execution environments.
Second, the results from early adopters are mixed. Retail investors using general-purpose models for trading report inconsistent outcomes, and the gap between agent-as-research-tool and agent-as-autonomous-trader remains the primary source of user risk.
Third, the legal and regulatory frameworks have not caught up. Stanford’s Phantom Agent paper provides a theoretical liability model, but no enforcement action, court ruling, or regulatory guidance has addressed an autonomous AI agent executing trades on behalf of a retail investor through a consumer brokerage.
Citizens’ Devin Ryan believes the majority of trades on some platforms will come from agents by end of 2027. If that projection holds, the regulatory and legal frameworks will be tested not in an academic paper but in customer complaints, enforcement actions, and potentially litigation. Financial services has a history of building products faster than regulators can evaluate them. Agentic trading is following that pattern.