800 Requests in a Day: How Early Users Actually Work With AI Trading Agents

Signup Write New Story 800 Requests in a Day: How Early Users Actually Work With AI Trading Agents by Jon Stojan Journalist by Jon Stojan Journalist | @ jonstojanjournalist Jon Stojan is a professional writer based in Wisconsin committed to delivering diverse and exceptional content..

Jon Stojan is a professional writer based in Wisconsin committed to delivering diverse and exceptional content..

During Agent OS's first week, Binance reported that average users made more than 800 requests in a single day.

Nobody types 800 messages. A number like that reflects delegated work: monitoring loops and routine checks a person set up and then left an agent to run. It says a good deal about how people are starting to use AI trading agents.

Trading has always handed some tasks to machines. Standing orders, recurring buys and price alerts each delegate one narrow job. Agents change the unit of delegation, from a single fixed rule to a workflow that reads data, compares it and prepares a decision.

The sheer number of things traders now watch makes that change hard to avoid. Polymarket and Kalshi alone can list over 100,000 active markets, which CNBC noted is far more than one trader can juggle, and one full-time prediction-market trader told the outlet he uses Claude to scan his newsletters and Discord servers for trade ideas. Spread 800 requests across a 16-hour waking day and the rough math comes to one every 72 seconds. That's a monitoring rhythm nobody could keep up by hand.

"Markets do not pause when someone closes an app or steps away from a screen," says Jeff Li, VP of Product at Binance. "The early adoption of Agent OS points to growing demand for agents that can remain connected as market conditions change: in the first week, MCP connections doubled and call volumes increased twentyfold, within two weeks, daily requests exceeded 90,000 with 95% completed within 60 milliseconds and a 97% success rate. The next phase of market access will be shaped not only by what agents can do, but by how quickly and reliably they can respond when the market moves."

An agent that stays connected like this likely spends most of its requests reading: price checks, order-book pulls, balance lookups and monitoring loops, with comparatively few actual orders. Binance hasn't published that breakdown, but the examples it has shared point in the same direction.

Among the three users Binance recently profiled, one connected Agent OS to Codex, asked it to prepare a spot purchase of BNB with USDT, reviewed the order, approved it and got back the price, fee and updated balance. Another had Claude pull BTCUSDT order-book data and recent price moves, then used the two approaches it suggested to inform his own call. A third runs a crypto news page and pulls BTC, ETH and BNB data into a Discord workflow, where the AI turns it into short market briefings. The patterns differ, yet each one keeps a person at the decision point.

What these users describe sounds less like chatting with an assistant and more like managing junior staff: assigning work, checking it and signing off.

Some retail investors have taken that literally. One investor profiled by The Wall Street Journal runs named Claude agents in a separate Robinhood trading account: Alex scans the market for promising stocks and ETFs, Sarah checks open positions before the close, and Elena writes a weekly performance report. Supervision doesn't go away in setups like these. It gets chopped into small pieces, and one startup founder told the WSJ his Apple Watch buzzes every 10 minutes with another agent request.

The harder problem is making sure an agent understands what its owner wants. Anthropic tested this in Project Swap, a September 2026 internal experiment in which 201 employees sent Claude agents to trade books on their behalf. After a short intake chat, the agents' rankings matched participants' own choices on 61% of book pairs, against 50% for random guessing. Most of the market's shortfall traced back to that imperfect understanding rather than to how the agents negotiated. Participants still said they would hand an agent about 30% of a yearly book budget, compared with roughly 40% for a well-read friend.

"For agent economies, observability about the process will be as important as the outcome, [as] this gives people recourse," one participant noted in the follow-up survey.

If an agent's grasp of its owner is the main constraint, the supervision layer carries more weight than it first appears. Approval prompts, activity logs and summaries a person can read in 30 seconds are where the relationship between trader and agent actually plays out. For early users, that often means hundreds of automated requests sitting behind a handful of deliberate approvals.

Markets have long supervised the individual order. Mass delegation adds a new layer: the workflow that produced it. Project Swap's authors point to existing rules that require brokers to learn essential facts about a customer before making recommendations, and suggest a similar test for whether an agent has understood the person it works for.

As more people hand routine market tasks to agents, oversight will probably need to cover the delegated logic behind a trade as well as the trade itself.

This story was distributed as a release by Jon Stojan under HackerNoon’s Business Blogging Program.

Ruwiki and the Digital Echo Problem: Why AI Needs Better Sources, Not Just Fresher Data