It sent an email
On 29 September Robinhood announced AI agents inside its app that can analyse the market, build a strategy and trade for you around the clock. The release is titled "Robinhood Puts the Power of Hedge Funds in Every Trader's Pocket." More than 150,000 agent accounts are already open, and the agents use its tools almost 30 million times a day. That is about 200 a day per account.
I built one of these in January 2020, with a partner, for our own money. In April 2021 we wrote the spec for its second version. The meeting notes are dated a Sunday, a Saturday and a Sunday. So I read the release against our spec.
The spec is a few pages. The trading signal takes one line of it. The step that ranks the stocks is labelled, twice, as my partner's black magic. The rest is schedule and bookkeeping: fundamentals once a week, parameters once a month, Brazilian stocks every hour, crypto every thirty minutes, performance against the index, a monthly report.
And one sentence about what comes out. Alerts will be sent by email.
The bot never placed an order. It looked, it decided, it wrote to us, and a person placed the trade.
The reason was Renaissance. I knew I could not compete with them. Two people with day jobs do not win at machine speed against a firm built for it, so I did not enter that game. I had to create my own methods, and the first one was the clock. If a signal is only good for a second, I will never get it. If it is still good an hour later, a person has time to read an email and place the order. We only looked for the second kind.
So the human in the loop was not a safety feature added to the bot. It was the statement of which edge we thought we had.
Back to the release, and to be fair to it. Each trade needs your approval unless you turn that off, and the setting starts on. The agent can only reach the money in its own account. The part that trades without asking is a feature announced as coming soon, for standing instructions that run overnight.
So at Robinhood the approval is a switch. That is a reasonable way to build it. I am not saying the product is wrong, and I have not used it.
I am saying the switch and our email answer different questions.
The switch asks how much you trust the agent. The email asked which game we were in.
A strategy that only works with the switch off is one where a person is too slow, and the firms in that headline are already there.
Our method had a price, and I paid it. The person in the loop was me, with a job. In the meeting notes my own item, finish testing the new method, is open on 11 April and still open on 30 May. And the spec has a function for adding what it calls stocks of the heart, the ones you hold whatever the model says. A person in the loop is slower than a machine and likes some stocks too much. I kept the person.
I am a vendor now, of an AI operations platform for industry. My market has its incumbents too, and I know what I knew in 2020: if I simply copy them, it is not going to work.
So the method is my own again. It starts where the bot's did and goes further. The bot looked, decided and wrote to a person. Our platform answers the question, keeps watch, and then acts on the data. It can raise the ticket, send the message, write the file, behind gates a person sets, and each action is recorded with its cost, its outcome and its undo.
What it never sends is a setpoint.
On the bot, the step that could not be taken back was the order, so a person kept it. In a plant that step is the setpoint, and a person keeps it.
Everything short of it is the software's job.
It is the same statement about edge. The bot's edge was a signal still good an hour later. Ours is work that waits a week: the question nobody asks because the answer takes the data team that long, the check that should run every day and is done by hand by people with other jobs. None of it is a race. It is a queue, and it gets shorter when the software does the work and the person is left with the decision.
The bot has run on my own money since January 2020. It has never placed an order.
Diego Mercadal started as a commissioning engineer on high-voltage motors and generators, joined an offshore drilling contractor as a rig hand, worked several positions in the drilling crew, and ended up running its AI/ML function. He is now co-founder and CEO of Wonder DataLabs.