Picture two AI trading agents running the same model and the same strategy. A signal arrives during a fast market, both send an order, and one platform answers in milliseconds while the other times out. Same intelligence, different result.

The difference sat in the plumbing. As agents take on more of the work of trading, the infrastructure connecting them to markets plays a growing role in what they can actually deliver.

The Constraint Is Moving From the Model to the Pipe For a consumer app, reliability is a user-experience problem. A delay means a spinning wheel, and the person taps again. For an agent it turns into a correctness problem. An agent making hundreds of automated calls reads a timeout as missing data, so it may act on stale account information, skip a check or retry an order into a worse price.

That gap is drawing money. Arcade raised $60 million in June on the view that most agents fail because the infrastructure around them isn’t ready, and TechCrunch reported that MCP servers have had to do extra work simply to track sessions across load-balanced machines, one reason large integrations were slow to ship.

“Markets do not pause when someone closes an app or steps away from a screen. 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,” says Jeff Li, VP of Product at Binance. “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.”Those figures reward a close reading.

A 95th-percentile result means one request in 20 took longer than 60 milliseconds, and fast markets are exactly where that slower tail matters. At a 97% success rate, a system handling more than 90,000 daily requests would see roughly 2,700 fail each day, by simple arithmetic. The published numbers also leave open where latency is measured and what counts as a failure.

Three Numbers That Compound at Agent Scale Latency, success rate and behavior under load look like separate metrics. For an agent they multiply. A person sends one request and reacts, while an agent chains several together. Suppose a workflow needs five sequential calls to check a balance and read the order book as well as size a position and place the order and confirm the fill, and each call succeeds 97% of the time independently. In that case the full sequence completes cleanly about 86% of the time.

Load is the hardest of the three, because demand spikes precisely when prices move and failures tend to cluster at the worst moment. MCP’s maintainers moved to a stateless design in July so servers can run behind standard load balancers without relying on one machine to remember each session.

“Before, you needed to have a session store and manage session IDs, and if one of your compute pods went down, all of a sudden the requests would start failing,” says Den Delimarsky, a lead maintainer of MCP.

Binance Research made a related point when it shared its two-week figures: as agents become more capable, the constraint may increasingly shift from model intelligence to the infrastructure needed to access data and execute reliably.

Automated Systems Fail at Machine Speed Trading has learned this before, at considerable cost. On August 1, 2012, a faulty software deployment at Knight Capital sent more than 4 million orders into the market in 45 minutes while trying to fill 212 customer orders. The firm traded more than 397 million shares and lost more than $460 million, according to the SEC, which later settled market access rule charges with Knight for $12 million.

For agent builders, the takeaway is uncomfortable. Automated systems can fail faster than anyone can step in, so reliability has to be designed in from the start.

Payments went through a similar evolution. Card networks and settlement systems still compete on availability and throughput, yet over time those properties became the price of entry for handling other people’s money. Agent platforms look set to follow. Once one publishes latency and success-rate figures, the others will be asked for theirs, and the published number turns into a floor.

Questions to Ask Before an Agent Touches Money Anyone connecting an agent to a funded account can put a few direct questions to the platform. At which percentile is latency reported, and is it measured at the server or end to end? What counts as a failed request, and what does the agent do after one?

Context matters as much as averages. How did the figures hold up during the busiest hour of the most volatile day, compared with a launch-week snapshot? Are there rate limits, and what happens to requests above them? Does the platform publish these numbers over time, or only once?

Agent platforms may increasingly publish execution metrics the way exchanges already publish system status, and the ones that do it consistently will make those questions much easier to answer. Contributed article. Not produced by the TNW newsroom and does not reflect the editorial stance of TNW.