For most of the past three years, progress in artificial intelligence has been scored the same way. A new frontier model arrives, a benchmark chart circulates, and the industry recalculates who is ahead. For the companies that actually ship software on top of those models, that scoreboard has quietly stopped explaining very much.
NxCode, an AI development platform that turns plain-English descriptions into working software, has raised a seven-figure U.S.-dollar investment from GSR Ventures. What the company says it will do with the money is the more interesting signal. The capital is going into the layer that sits above the model: The specialized agents that handle planning, architecture, coding, testing and deployment, and the system that decides whether their output should be trusted at all.
That emphasis reflects a change in what is actually scarce. Producing code has stopped being the constraint.
The AI coding boom has created a trust problem
Sonar’s 2026 State of Code Developer Survey, based on responses from more than 1,100 professional developers, estimated that AI now accounts for roughly 42 percent of committed code. The same study found that 96 percent of those developers do not fully trust what the machine produces, that fewer than half always verify it before committing, and that 38 percent find reviewing AI-written code more effortful than reviewing a human colleague’s.
The 2025 Stack Overflow Developer Survey pointed the same way. Distrust of AI accuracy outweighed trust among respondents, and 66 percent named output that is “almost right, but not quite” as their single largest frustration. A program that is obviously broken is cheap. A program that is plausibly wrong is not.
Better models won’t solve the judgment gap
This is the gap a better model does not close. Raising average suggestion quality does not tell anyone which suggestions matter, which failures are tolerable, or when a system should stop and ask a question rather than proceed. Those are not generation problems. They are judgment problems, and the survey data shows developers guarding them closely: 76 percent of Stack Overflow respondents said they had no plans to hand deployment and monitoring to AI, the stages at which a suggestion becomes an operational fact and accountability attaches to a person.
Zirong Chi, who founded NxCode and leads its product design, development and operations, puts the strategic case in economic terms.
“Anyone starting a software company has to think about how the world is going to change,” she said. “Whatever can become free will become free. What does not get commoditised is judgment.”
It is an uncomfortable prediction for a large part of the current market. If capability is a shared input that gets cheaper every quarter, then a thin interface onto someone else’s model has a short shelf life, and the durable position is whatever remains hard once generation is free - deciding what is worth building, recognising which errors are survivable, and being answerable when one is not.
Why NxCode is betting on AI agents and human trust
Chi’s route into the company was not the usual one. She was living a comfortable retirement in Japan before starting NxCode, and describes the decision as a reaction against the way problems tend to get discussed rather than solved.
“A lot of people’s way of thinking about a problem is not actually aimed at solving it,” she said. “I have real respect for anyone willing to face one head-on. That is most of why I stopped being retired.”
The instinct predates this company. Her previous product, Bibabo, taught AI and programming through debugging exercises rather than presenting model output as something to be accepted at face value - a design that treats interrogating a result, not obtaining one, as the skill worth building.
The stakes are highest for the audience this whole category has courted hardest. Tools that turn a description into an application are marketed at people without engineering backgrounds, and those users are the least equipped to notice the failures that do not announce themselves. A generated product can present a finished surface while something underneath is quietly wrong: Password recovery, a database with no safe migration path, an unusual payment case, or the boundary of what an agent is allowed to retain and repeat.
NxCode’s answer is structural rather than rhetorical. The platform splits the work across specialized agents, and the company describes the review layer as the surface it intends to compete on: What gets checked without being asked, which risks are explained in ordinary language, when the system declines to continue, and how a user can inspect, export or roll back what the agents produced. NxCode was selected for MiraclePlus’s F25 program and says more than 5,000 non-technical builders now use it.
The next AI advantage may be knowing when to stop
Whether any of that works is measurable, and the measures are unglamorous. What share of projects reach production. How many are still running six months later. How often a user has to fall back on human help. Which errors keep recurring. Whether a customer can leave with their code intact. None of those numbers appear on a benchmark chart, and none of them improve simply because the underlying model did.
The model race will carry on, and the leaderboard will keep changing hands. But for the companies building on top of it, the competitive question has moved. Capability is becoming an input anyone can buy. What stays scarce - and what this round is meant to buy time to build - is a system a non-engineer can reasonably rely on, and a clear answer to who is responsible when it turns out to be wrong.
NxCode, an AI development platform for non-technical builders, is incorporated in Delaware and operates from Mountain View, California, in the United States.
This article was written in cooperation with NxCode