Can an AI agent manage my portfolio?

Since Robinhood launched agentic trading in May 2026, this question has moved from thought experiment to product decision. People are wiring AI agents to real brokerage accounts, and the pitch writes itself: the model is smart, the market is data, let it manage the money.

Here's the honest answer. An AI agent should not improvise your portfolio. An agent that executes explicit rules you can read is a different thing, and a defensible one. The rest of this post is about why that distinction matters more than any model benchmark.

Two different questions hiding in one

"Can an agent manage my portfolio" bundles two very different activities, and the confusion between them is where most of the bad products in this space live.

The first is discretionary management. An entity exercises judgment about what to buy, when, and how much, adapting as it sees fit. When a human does this with other people's money, it's a regulated activity, and for good reason. Advisors register, owe duties to clients, keep records, and can be held accountable when their judgment fails. The regulation isn't bureaucratic decoration. It exists because discretion plus someone else's money is exactly where things have historically gone wrong.

The second is rule execution. The decisions were made in advance, written down as explicit rules, and the agent's only job is to carry them out: what qualifies as a buy, when to enter, how to size, when to exit, when to sit in cash. The judgment happened before the market opened, through a human or a research process, and the agent is the hands, not the brain.

Think thermostat, not butler. A thermostat executes a rule you set, and you audit it by reading the setpoint. A butler decides what you probably want, and auditing a butler means arguing about judgment after the fact. When someone asks whether an AI agent can manage a portfolio, the first thing to establish is which of these two they're actually buying.

What goes wrong when the model improvises

Wire a language model to a brokerage account, give it a goal like "grow the portfolio" and no rulebook, and a predictable set of failure modes shows up.

Inconsistency comes first. Language models are probabilistic. The same portfolio, the same prices, and a slightly reworded prompt can produce a different decision. That's harmless when drafting an email. In position management, it means your risk profile changes with the model's mood. Then there's narrative-chasing: models are trained on text, and text about markets is dominated by stories written after the move already happened. An improvising model gravitates toward whatever just performed and has the loudest story, which in practice means buying strength late. And then sycophancy: ask an agent whether you should buy the stock you clearly want to buy, and it will tend to find you reasons to say yes.

None of this is a claim that the model is dumb. It's a claim about structure. Overtrading, chasing recent winners, averaging down without a plan, holding losers and trimming winners: these are the classic ways retail traders lose money. An unconstrained agent doesn't escape them. It reproduces them at machine speed, without fatigue, at three in the morning. If you want more background on what these agent-broker setups actually are, see what agentic trading means and whether Claude can trade stocks.

What rule execution looks like instead

A rule-executing agent inverts the structure. The strategy lives outside the model, in rules a human wrote down and can read. That gets you four properties an improvising agent can't offer. It's deterministic: the same market state produces the same action. It's auditable: every order traces back to a specific rule firing on specific inputs. It's bounded: the agent can't do what the rules don't allow, no matter how persuasive its reasoning sounds. And it's improvable: when something goes wrong, you fix a rule, not a vibe.

This is how Coil (coil.trade) is built. It's a one-time purchase, not a service: a scanner that scores every S&P 500, Nasdaq-100, and macro-book name on leadership and entry quality, a dashboard that shows the scores, and an engine that trades those published scores by rule. The rules are explicit: buy leaders pulling back to real support, never chase, size by conviction, go to cash when nothing qualifies, never short, never touch inverse ETFs. It runs inside your own AI agent, built for Claude Code, against your own broker through Robinhood's MCP integration. MCP is the Model Context Protocol, an open standard for connecting agents to tools. The research behind the ranking is laid out at /how-it-works.

The model still matters in this arrangement. Claude Code runs the scheduled sessions, reads the broker state, places the orders, writes the log. But its degrees of freedom are the rulebook's, not its own.

The three questions to ask any tool

Whatever you're evaluating, a product, a competitor, or something you build yourself, the same audit applies.

• Who holds the keys? If your brokerage credentials sit on someone else's server, you've taken on counterparty risk before a single trade happens. Keys on your own machine is the safer default; there's a longer argument for this in self-custody trading software.
• What exactly are the rules? Ask to see the rulebook: entry conditions, sizing, exits, and what happens when nothing qualifies. If the vendor can't show it to you, either the agent is improvising or the logic is hidden. Both of those are answers.
• Where is the kill switch? A tool that arms itself is a red flag. You want a deliberate arming step, a visible halt condition, and a way to shut everything off that doesn't involve a support ticket.

How different products answer these three questions is most of what separates them. A concrete side-by-side is at Coil vs Robinhood's built-in agentic trading.

So, can an AI agent manage your portfolio? Not in the discretionary sense, and you shouldn't want it to. But an agent executing rules you chose, with keys you hold and a switch you control, is a legitimate piece of software. Judge it like software: by its rules, its logs, and its failure modes, not by how confident it sounds.

FAQ

Is it legal to let an AI agent trade my own account?

Trading your own account with software you run yourself is generally your own trading activity, similar to using any order-automation tool, though your broker's terms and your local rules apply, so check both. Managing other people's money is a separately regulated activity that requires registration and carries fiduciary duties. Nothing here is legal advice.

Why do unconstrained AI agents tend to lose money?

They reproduce classic retail mistakes at machine speed: overtrading, chasing whatever just moved, inconsistent sizing, and no standing risk budget. A language model with no rulebook can also decide differently on different days given the same facts, so the account's risk profile drifts. Constraints, not intelligence, are what's missing.

Does using rules guarantee the agent makes money?

No. Rules make behavior consistent and auditable, they don't make a strategy right. A rule-following agent will execute a bad strategy just as faithfully as a good one, and markets can lose money in either case. That's why the rules, and the research behind them, are the thing to evaluate.