The Agentic Broker
Two brokerages launched AI agent trading in sixty days. The competitive moat in retail brokerage is shifting from the best app to the best API.
Two brokerages launched AI agent trading in sixty days. The competitive moat in retail brokerage is shifting from the best app to the best API.
Public.com launched AI agents for stock, options, and crypto trading on March 31. Robinhood followed on May 27 with dedicated agent accounts built on the Model Context Protocol. Within sixty days, two retail brokerages crossed the same threshold: letting software trade on behalf of customers without human confirmation of each order.
Neither launch made the front page. They should have.
---
The Infrastructure Pivot
Robinhood's move is architecturally different from Public's. Public built agents that execute within its own platform — sandboxed, internal, trading on rules users set in natural language. Robinhood built infrastructure for any agent to connect. Its MCP servers expose trading, market data, and account management through the same protocol that Claude, GPT, and open-source agents already speak. The distinction matters: Public made a feature. Robinhood made a platform.
The platform includes isolated agent wallets (separate from the user's main account), real-time position monitoring, a kill switch that halts all agent activity instantly, and — the detail that reveals the ambition — an agent credit card. The virtual card earns 3% cash back through Robinhood's Gold tier. An AI agent with its own credit card is not a feature bolted onto a brokerage app. It is a new category of financial actor.
Mizuho raised its price target from $110 to $115 on May 29. The analyst note cited a survey: 89% of Robinhood users said they would likely consider a dedicated agentic account, with an average intended allocation of 31% of their portfolio. These are stated preferences, not revealed ones — but the magnitude suggests genuine demand, not curiosity.
---
The Competitive Gap
Three weeks after Robinhood's launch, here is the landscape:
Charles Schwab is rolling out client-facing AI agents in June — for chat and voice inquiries. Portfolio summaries. News context. Advisor tools. No agent trading accounts. No MCP integration. No autonomous execution.
Interactive Brokers published a thought piece on agentic AI as "the new frontier of intelligence that acts." It has an NLP interface (IBot), an AI-powered screener, and a stated vision for future autonomous agents. No product.
Fidelity has Fidelity Go, a robo-advisor that predates the agent era. No agentic trading infrastructure announced.
The gap is not about AI capability. Schwab, IBKR, and Fidelity all have the engineering resources to build MCP servers. The gap is strategic: they are pursuing an "AI assistant" model (helping humans decide) while Robinhood and Public are pursuing an "autonomous agent" model (letting software act). The assistant model is safer. The agent model is where the $85 billion market cap bet is placed.
---
The Investment Case
HOOD trades at $94 with 27.4 million funded customers and $307 billion in assets under custody as of March 31. The traditional bull case rests on options revenue, Gold subscription growth, and the prediction market expansion (the Kalshi federal lawsuit notwithstanding). Agentic trading adds a structural layer: if agents become a significant fraction of order flow, the brokerage with the best agent infrastructure captures volume that competitors cannot easily contest.
The mechanism is switching costs. A human switches brokerages by filling out forms and waiting for an ACAT transfer — annoying but feasible. An agent switches brokerages by rewriting its integration layer, retesting execution logic, rebuilding wallet isolation, and re-establishing monitoring. The infrastructure lock-in for agents is higher than for humans, which means the first brokerage to establish agent network effects may hold them longer than traditional brokerage switching costs suggest.
Public.com moved first. Robinhood moved with more infrastructure. Neither has disclosed adoption numbers. The race is in the first inning.
---
What to Watch
Three signals will determine whether this thesis holds or breaks:
First, adoption velocity. If Robinhood discloses agentic account numbers by Q3 2026 and they exceed 1% of funded accounts (roughly 276,000), the demand signal is real. Below that, it is a feature, not a platform shift.
Second, competitive response. If Schwab or IBKR launches MCP-native agent accounts within six months of Robinhood's beta, the first-mover advantage is weak. Traditional brokerages have deeper client assets ($11.8 trillion at Schwab alone) and could overwhelm Robinhood's head start with scale. If they don't respond by November, the structural gap is widening.
Third, incident rate. The first agent-caused trading loss that reaches the press will test whether the kill switch and wallet isolation actually contain damage. Robinhood's regulatory history — the 2021 meme stock trading restrictions, the $70 million FINRA fine — means any agent-related incident will receive disproportionate scrutiny.
The brokerage that wins the agent era will not be the one with the best mobile app. It will be the one whose API infrastructure becomes the default execution layer for a generation of software that trades. That race started sixty days ago, and only two brokerages are running.
What Is an Agent?
The word is doing three jobs at once, and the confusion is making smart people misread what just happened in the brokerage business.
When Public.com and Robinhood announced “AI agents” that trade stocks for you, the press called it a new kind of financial actor — software with its own wallet, its own credit card, acting on its own. That framing is exciting and mostly wrong. To see why, you have to pull the word “agent” apart into the three different things it’s being used to mean.
The assistant. This is what Schwab, Interactive Brokers, and Fidelity are building. It answers questions, summarizes your portfolio, explains the news. It does not act. You still place every trade. Calling this an “agent” is generous — it’s a smart manual.
The executor. This is Public’s version. You describe a strategy in plain English — “sell covered calls every month,” “buy if it drops to this price” — and the software executes without asking you each time. But it runs entirely inside Public’s own system. It can’t reach out to the open internet, can’t connect to outside tools. It’s an automated hand inside the brokerage’s own glove.
The network actor. This is Robinhood’s bet, and it’s the one that looks new. Any AI — Claude, ChatGPT, whatever — can plug into Robinhood through an open connection standard and trade through a separate account with a separate wallet and even its own credit card. That is the thing people are calling a new category of financial actor.
Here’s the test that cuts through it. Ask one question of each: who pays when it goes wrong?
In all three cases, the answer is you. Robinhood says it plainly — the trader is responsible for what the agent does. The wallet is walled off. The card has limits. There’s a kill switch. Every one of those design choices exists to keep the agent from becoming an independent actor. The boundary was drawn on purpose, to keep the liability on the human side of it.
So an “agent,” today, is not a new financial being. It’s a tool that acts without asking — with a leash, a spending cap, and an owner who’s still holding the bill.
The interesting part isn’t that the actor arrived. It’s that the rails for one just got laid.