Should I let an AI agent trade my stocks?

Letting an AI agent trade your stocks: is it worth the risk?

Probably not — at least not with full discretion. Robinhood's May 2026 MCP launch lets AI agents place real trades, and the company itself warns of "the possible loss of your entire investment." The right use, as Kahneman's work on overconfidence implies, is to let agents analyze, alert and propose, but keep the final click human.

I am genuinely excited about agentic AI, and I still do not let an agent place trades on my behalf without a human approval step. The Robinhood announcement on May 27, 2026 — connecting AI agents to brokerage accounts through their Model Context Protocol service so they can analyze concentration risk, screen markets, and execute trades — is one of the more interesting pieces of plumbing to ship this year. It is also exactly the kind of capability where the gap between "I can do this" and "I should do this" is widest.

Let me lay out what the integration actually does, because the headlines blur it. Through MCP, your AI agent can read your positions, look at sector exposure, scan options chains, and place orders. Robinhood added safeguards — required confirmations for certain actions, spending caps, the usual disclaimers — and explicitly warned in its own announcement that AI-powered trading "involves significant risk, including the possible loss of your entire investment." That is not lawyer-speak you can wave away. That is the operator of the platform telling you, in writing, where the failure mode lives.

The reason I will not give an agent full discretion has nothing to do with whether the model is smart. It has to do with what Daniel Kahneman taught us in Thinking, Fast and Slow about overconfidence and the illusion of validity. Markets are one of the cleanest natural experiments we have for testing whether smart, confident agents — humans or otherwise — can actually predict short-term outcomes. The answer is consistently no. Stock pickers, Kahneman pointed out, show zero year-to-year correlation in their performance. Expert political forecasters do no better than dart-throwing chimps. Subjective confidence, he wrote, "is a feeling, not a judgment." An AI agent will sound extraordinarily confident about a thesis it generated in eleven seconds. That confidence is a feeling too. It tells you nothing about whether the trade is sound.

There is also a specific failure pattern I keep watching for in my own life and in the founders I coach: the moment you fully delegate something you care about, you stop paying attention to it. This is true for hiring, for content, and it will be especially true for money. If an agent runs my portfolio, I will check on it less often than I check now. When something goes wrong — a flash event, a regime change, the agent misinterpreting a structural shift as noise — I will be a step behind. Robert Iger in The Ride of a Lifetime describes the discipline he had to keep as CEO of Disney: even when he had brilliant lieutenants, he stayed close enough to the decisions that mattered to be the one accountable for them. Money is one of those decisions for most of us.

What I do think is genuinely useful is the analytical half of what MCP unlocks. Letting an agent pull your full position data and run an honest concentration-risk analysis, or flag that your exposure to one sector quietly drifted from 12% to 31% over the last quarter, or watch for the kind of patterns you would never spot scrolling through the app — that is high-leverage. So is having an agent draft a tax-loss-harvesting plan you then approve trade by trade, or pre-screen candidates and explain its reasoning before you click. The agent earns its keep as the analyst and the second opinion. You stay the principal.

My one-rule policy is simple: any trade above a number I set in advance — for me, anything bigger than a routine rebalance — has to be proposed by the agent and clicked by me, with the agent's reasoning in front of me when I do it. That preserves the speed advantage where it actually matters (catching things I would miss) without giving up the accountability where it actually matters (the trade itself). It also keeps me honest, because reading the agent's case forces me to engage System 2 instead of nodding along.

The deeper point is the one Naval Ravikant makes about leverage in his Almanack: leverage amplifies judgment, in both directions. AI agents are the most powerful leverage tool we have ever had. If your judgment is sound, agents make you faster and sharper. If your judgment is off — or if you outsource it entirely — they will lose your money faster than you ever could on your own. Use the leverage. Keep the judgment.