Robinhood co-founder and CEO Vlad Tenev arrived for the conversation short on sleep. "Last night was rough with my coding agents," he told the hosts of Moonshots with Peter Diamandis, in an episode recorded on September 18, 2026 and published the next day. Dave Blundin, founder and general partner of Link Ventures, asked whether the agents wake him up. Tenev said he lets them run and tries to sleep with the technology in another room, partly out of concern for his mental health.
Blundin described his own drift toward looser control. He originally set spending budgets for his AI agents through the software consoles, he said. Now he simply tells them something like "spend a thousand dollars on that but no more", walks away and assumes they will comply. "There's going to be like a seven figure bill one of these mornings," he joked.
Later, Peter Diamandis, the founder of XPRIZE and Singularity University, turned that question of delegated control to money. With thousands of agents trading on Robinhood's platform, he asked Tenev, what protections and structure had the company put in place?
A separate account, with a small starting balance
Robinhood's offering is called Agentic Trading. An AI agent here means a model such as Anthropic's Claude that can take actions through software tools, not just answer questions. Tenev said the product starts with a separate brokerage account, segregated from a customer's main Robinhood account and retirement account. Customers have to create the agentic account and deliberately move money into it. He said people typically fund it with $100.
The company began with stock trading only, with no leverage and no margin (borrowed money), because it wanted to learn what people would do with it. Tenev described that first version as "fairly cabined in". Robinhood has since added options trading, limited margin and, more recently, crypto, which he said it has started rolling out.
The separation covers where the agent can trade, not everything it can see. Robinhood's Agentic Trading documentation says agents connect through the Model Context Protocol (MCP), a standard interface that lets AI assistants look up information and take actions in other services. Trading is confined to the dedicated account. However, the documented read permissions extend across a customer's Robinhood accounts, covering account numbers, positions, balances, transaction histories, watchlists and scans. When a customer authorizes it, an agent can place trades without asking for confirmation of each one. The documentation warns that agents can misunderstand instructions, act on stale information and lose the entire amount invested. For crypto, agents can trade eligible pairs, but customers must handle transfers, staking or lending themselves.
Hard to set up, and models that don't want to trade
Tenev said two early lessons stood out. First, customers currently have to do all of their trading from inside Claude Code or OpenAI's Codex, which are coding tools. He noted that many people in the hosts' circle use Claude Code, but very few members of the general public do. Connecting such a tool to a service like Robinhood is "pretty complicated for a lot of people," he said, and the company has been thinking about how to simplify that step.
Second, because Robinhood does not control the underlying model, the models often balk. Asked to deploy a trading strategy, Tenev said, they will respond along the lines of "I don't really feel like trading right now."
Later in the episode, Diamandis said Robinhood had more than 100,000 accounts running AI agents for trading. He said it was his understanding that the agents often refuse not for risk reasons but because trading examples are missing from the data the models were trained on, and asked whether that was the case. Tenev agreed that the models are "just not trained for that," and added a second factor: requests may resemble activities that model developers have explicitly tried to guard against. He said the models are getting better. The refusals still show how much work remains, he said, not only on the user interface and brokerage infrastructure but also on getting models to use Robinhood's tools more effectively.
Tenev expects more specialized models in future. He said companies are likely to deploy them in-house, either by fine-tuning existing open-weight models, whose parameters are publicly downloadable, or by training their own for particular uses. The goal would be models that become very good with each company's tools and data. He said he had heard that argument as theory, set against the view that general models keep improving. Now Robinhood is experiencing directly how hard it is for one model to be good at many specialized tasks, he said.
"Where's the alpha?"
Alexander Wissner-Gross, a computer scientist and founder of Reified, raised what he called "the obvious question". In investing, "alpha" means returns above what the market as a whole delivers. Wissner-Gross said quant funds, which trade using mathematical models and computers, compete viciously, and that the average lifetime of a new quant fund or strategy is short. Algorithmic traders already dominate trading volume, he said, and individual day traders have a difficult time; he guessed they generally cannot find alpha. So why, he asked, should a day trader delegating to Claude or another model, without the latency and other advantages of a large quant fund, expect any alpha at all?
Tenev did not claim that the agents beat the market. Right now, he said, what Robinhood sees is "a lot of automation type use cases." His example was an iron condor, an options strategy assembled from several separate contracts, or "legs". Putting one together takes a lot of manual work: visiting different websites, checking signals, constructing the trade and deploying it on a regular basis. The trader still supplies the idea, Tenev said, while the agent helps with assembly and execution, "removing the paper cuts."
Giving everyone the technology of a highly sophisticated quant fund, he said, is "a huge roadmap" and a moving target because such firms keep improving. It would require more and higher-quality data, better intelligence, very good deterministic code (software that follows fixed rules) and advances in latency and performance. The most sophisticated firms, he noted, compete over who gets "port one on the switch in the data center." Robinhood is at the beginning, he said, "going into the fog" while building the product and its infrastructure at the same time.
The advantage Tenev claimed is breadth. Robinhood spreads the cost of its market connections across countries and asset classes, he said, offering stocks, options, futures, prediction markets and on-chain assets on its own blockchain. As it expands, he argued, the platform will have advantages over what at least a startup quant fund could connect to.
Rules written by AI, or an AI making the call
Earlier in the conversation, Blundin drew a distinction he said many people miss. AI can be used to write automated code that behaves deterministically, or it can sit inside the decision loop, choosing what to do as events unfold. Those are "very, very different things," he said. A trader's own stop-loss (an order to sell if a price falls to a set level) or a planned trade around world events can easily become fixed, executable code. The temptation is to put the AI in the loop for judgment calls, such as whether it is a turbulent day or whether there is trouble in the Middle East, because that makes the system so much easier to build. But then, Blundin said, a third party's decisions sit inside your own decision-making, and it is not clear whether that agent bears any liability.
Plumbing or fiduciary?
Wissner-Gross pressed the point about responsibility. It sounded, he said, as though Robinhood was agnostic about whether users achieve alpha: if they lose a lot of money day trading or delegating to their algorithms, that is not its problem. "No crying in the casino," he said; Robinhood is "just the plumbing."
Tenev said that was "basically true" for active-trading products, with a caveat. Active traders know what they want to do, and Robinhood's job is to be a tool provider. It focuses on execution and plumbing, he said, but also wants to offer analysis and research tools, data sets and models that could help traders develop better strategies.
For customers who do not want to make trading decisions, he pointed to a different category: products under the Robinhood Strategies umbrella, where the company acts as a fiduciary, obliged to act in the client's interest. A customer deposits money and Robinhood manages it. One option is a Smart Income portfolio geared toward generating yield. Tenev described a slider on which customers set a target yield for each level of risk. Robinhood's Smart Income page organizes the managed portfolios around time horizons instead. Shorter horizons lean on Treasuries and high-quality bonds, while longer ones add dividend stocks and bond ETFs and accept more short-term price movement. Robinhood Strategies picks the investments and handles rebalancing, and the page warns that principal can be lost and that displayed yields are illustrations, not guarantees.
Tenev also mentioned a third route. Through its acquisition of TradePMR, Robinhood can connect customers with human advisers for portfolio management, estate planning and taxes. The agentic accounts sit at the do-it-yourself end of that range: the trader decides and the agent executes. For now, by Tenev's account, what the agent mainly saves is effort.