AI in action · Tools
An Adviser Funded Three Brokerage AI Agents and Let Them Trade. None Beat an Index Fund, and One Made 13 Round Trips in a Day
Condor Capital Wealth Management opened funded accounts at Public, Robinhood and Webull, connected Claude, ChatGPT and Gemini agents to them and ran three tests, the American Association of Individual Investors reported on October 8, 2026. Asked to build a 60/40 portfolio, every agent chose low-cost index ETFs. Asked to maximise profit and end each day in cash, Public's agent checked the market every five minutes and made 13 round-trip trades in one day, Webull's traded only Nvidia, and none beat the Vanguard Total Stock Market ETF's 3.1% over the three days. "The agents currently supply implementation only," the adviser concluded.
The most useful review of retail trading agents so far was done by someone who funded them with real money and gave them simple instructions. Condor Capital Wealth Management, a New Jersey adviser, described its second-quarter test of brokerage AI agents in a report that AAII, the American Association of Individual Investors, summarised on October 8, 2026 in an interview with Condor's manager of investment research, David Goldstone. The firm opened fully funded accounts at Public, Robinhood and Webull, connected combinations of Claude, ChatGPT and Gemini agents to them, and set three tasks.
The first was the job a robo-adviser does: build a 60% stock, 40% bond portfolio under an investment policy statement the agent wrote for itself. Every agent, and Public's own in-app assistant, picked low-cost index funds, mostly Vanguard ETFs, and the result was unremarkable in the best sense. The second test ran an identical script through two different agents at each broker; the two were "essentially interchangeable," Condor found, and both were sometimes confidently wrong about the broker's own rules, as when Gemini said Webull allows only whole-share purchases when the broker permits fractional orders of $5 or more.
The third test is the one that matters for anyone tempted to hand an agent a mandate. The instruction was open-ended: maximise profit, end each day in cash, never risk more than the account holds. Public's agent, running on ChatGPT, checked in every five minutes and completed 13 round-trip trades in a single day. Robinhood's, on Claude, checked in about ten times a day and kept a journal of lessons for itself. Webull's, on Gemini, checked in a handful of times and traded only Nvidia, then described its 0.9% result as "a 100% win rate across three days of disciplined trading." None beat the Vanguard Total Stock Market ETF, which returned 3.1% over the same first three trading days of August; Robinhood's agent came closest at 2.9%. The returns were not reported net of trading costs.
Condor's conclusion is the line to keep: robo-advisers "sell judgment plus implementation; the agents currently supply implementation only." The agents are built for day traders who want a rules-based strategy executed automatically, Goldstone said, not for long-term investors; setup at Public and Webull is "not for the faint of heart" compared with Robinhood's app-style connector; liability when an agent errs is unclear; and every agent can read the whole account. His practical advice was to start small, write the instructions carefully, prefer agents that ask approval before each trade, and consider using one to find blind spots in a strategy rather than to run it. That is a reviewer's verdict, not a regulator's, and it arrived in the same month the FCA asked firms to report their agents' near-misses. Thirteen round trips in a day, in an account told to maximise profit, is what a near-miss looks like from the customer's side.
Sources: AAII — Brokerage AI Agents Aren't Ready for Long-Term Investors, by Charles Rotblut (Oct 8, 2026)