70,000 accounts out of 25 million: how early agentic trading really is

At its July 2026 keynote, Robinhood said that roughly 70,000 accounts had enabled agentic trading. The feature launched in May 2026, so that is about two months of adoption, announced on the company's own stage, in its own words.

Robinhood also has more than 25 million funded accounts. Put the two numbers side by side and the real story appears: well under 1 percent of the customer base has turned the feature on. For all the noise around AI agents this year, almost nobody is letting one touch a brokerage account yet. If the term itself is new to you, start with our primer on what agentic trading is, then come back. This piece is about what the 70,000 figure actually tells us.

Under 1 percent is what a beginning looks like

It is tempting to read 70,000 as small. It is small. But small is what the front edge of an adoption curve looks like, and the shape of the curve matters more than the point you happen to be standing on. Online brokerage accounts were once a rounding error next to phone orders. Index funds spent years as an academic curiosity before they became the default. In both cases the early number was easy to dismiss, and the direction was not.

Two months in, the honest description of agentic trading is launched, real, and barely adopted. That is not a knock on Robinhood. A brokerage letting software place orders on a customer's behalf should want adoption to be slow and deliberate. The interesting question is not why the number is small. It is what has to be true for the number to get big, and whether those things are happening.

Who the first 70,000 probably are

Robinhood has not published a breakdown, so treat this as inference rather than fact. But early adopters of technical products look fairly consistent, and agentic trading has a steeper on-ramp than most consumer features. To use it today you need to be comfortable running an AI agent, connecting it to a broker over MCP, the open standard for agent-tool connections, and granting software the authority to act instead of just answer.

That points to a specific crowd: developers who already run agents like Claude Code on schedules, quant-curious traders who wanted automation without building the plumbing themselves, and tinkerers who enable new features on principle. Most are probably experimenting with small balances, which is exactly right. If you want to see what the wiring involves, our guide to connecting an AI agent to Robinhood walks through it end to end.

What the first 70,000 are not, in all likelihood, is representative. The next million adopters will be less technical, less careful, and easier to sell to. That is where the infrastructure question stops being academic.

The plumbing that still has to mature

For agentic trading to move from under 1 percent to something mainstream, a lot of unglamorous infrastructure has to exist first. The pieces are visible but unfinished:

• Authorization standards. Today each broker and each agent framework handles permissions its own way. An agent-checkout standard, a common grammar for what an agent may do, spend, and touch, would let people grant narrow authority instead of handing over the whole account.
• Safety norms. Sensible defaults are still a differentiator when they should be table stakes: software that ships disarmed, position limits, kill switches, and human confirmation for anything irreversible.
• Track records. There is no standard way yet for an agent system to prove how it has actually traded. Until verifiable records exist, marketing fills the gap, and marketing in this niche has a poor history. Our notes on AI trading scam red flags exist for a reason.
• Broker-side rails. Rate limits, scoped credentials, and order-review steps designed for software callers rather than adapted from human flows.

None of this is exotic. All of it takes time, and most of it will get built in response to failures rather than ahead of them.

Early does not mean profitable

Here is the part that adoption-curve enthusiasm tends to skip. Enabling an agent is a settings change. Trading well is not. An agent with market access and no tested rules inherits every human failure mode at machine speed: overtrading, chasing strength, averaging into losers, holding with no exit plan. Most naive setups, meaning a language model pointed at a brokerage with a vague instruction to make money, will lose it. Slowly through churn, or quickly through one bad concentrated position.

This is why the interesting systems in this space are rulebooks first and agents second. Coil (coil.trade) is built on that premise: the agent executes a published, inspectable process, buying leaders pulling back to real support, sizing by conviction, and sitting in cash when nothing qualifies, rather than improvising in a chat loop. The research behind the ranking is laid out at /how-it-works. Whether that specific process suits you is your call, and no process removes the risk of loss. But the gap between an agent with a rulebook and an agent with a vibe is the gap worth watching as this era grows up.

What to watch next

A few markers will tell you how fast this is actually moving. Watch the next number Robinhood reports, because the growth rate matters more than the level, and 70,000 is now the baseline. Watch whether other brokers ship agent interfaces of their own; one broker with an MCP server is a feature, several is a category. Watch for authorization standards that get real adoption instead of staying proposals. And watch for the first credible, verifiable agent track records, because the first team that can prove its results, good or bad, will reset expectations for everyone else.

If you are weighing Robinhood's built-in agentic mode against a dedicated system, we wrote a straight comparison at Coil vs Robinhood agentic trading. The safest read of the keynote number is this: the era is real, it is very early, and the people who do well in it will be the ones who treat an agent as software to be governed, not a genie to be prompted.

FAQ

How many Robinhood accounts have enabled agentic trading?

At its July 2026 keynote, Robinhood said roughly 70,000 accounts had enabled agentic trading. With more than 25 million funded accounts, that is well under 1 percent adoption.

Does early adoption mean agentic trading is profitable?

No. Enabling an agent is a settings change, not a strategy. Most naive setups, meaning an agent with broker access and no tested rules, lose money through overtrading and poor exits. Any experiment should use strict limits and money you can afford to lose.

What should I check before letting an AI agent trade for me?

Look for a written, inspectable rulebook, software that ships disarmed so arming is a deliberate human step, credentials that stay on your machine, hard position and risk limits, and a log of every decision. That checklist applies to any tool, including Coil.