converted
However, despite these issues, there is probably no way to look back. “The prospect of having to write every line of code by hand on a keyboard — those days are rapidly slipping behind us,” says Kyle Daigle, chief operating officer of Microsoft-owned code-hosting platform GitHub.
The StackOverflow report found that despite growing distrust in the technology, its use has grown rapidly and steadily over the past three years. Erin YepisA senior analyst at Stack Overflow says this shows that engineers are taking advantage of the tools with a clear view of the risks. The report also found that frequent users are more enthusiastic and that more than half of developers are not using the latest coding agents, perhaps explaining why many people are overwhelmed by the technology.
Those latest tools may be a revelation. Trevor DalyThe CTO of software development agency Twenty20 Ideas says he found some value in the autocomplete functions of AI editors, but when he tried something more complex it “failed catastrophically.” Then in March, while on vacation with his family, he set up the recently released Cloud Code to work on one of his hobby projects. It completed a four-hour task in two minutes, and the code was better than the code he had written.
“I was like, wow,” he says. “That, for me, was really that moment. There’s no going back from here.” Dilly has since co-founded a startup god herdWhich is creating software that can marshal multiple agents to work in parallel on a single piece of software.
Challenge, say Armin Ronacher, A leading open-source developer says the learning curve for these tools is shallow but long. He wasn’t impressed with AI tools until March, but after leaving his job at software company Sentry in April to start a startup, he started experimenting with agents. “I spent several months basically doing nothing else,” he says. “Now, 90% of the code I write is AI-generated.”
Getting to that point requires extensive trial and error to figure out which problems might trouble the devices and which they can handle efficiently. Today’s models can tackle most coding tasks with the right guardrails, says Ronacher, but these can be very task and project specific.
To get the most out of these tools, developers must give up control over individual lines of code and focus on the overall software architecture. Nico WesterdaleChief Technology Officer at veterinary staffing company IndeVets. He recently built a data science platform consisting of 100,000 lines of code almost exclusively by inspiring models rather than writing the code himself.
Westerdale’s process begins with an extended conversation with ModelAgent to develop a detailed plan for what to build and how. Then he guides him through each step. Westerdale says things rarely get done right on the first try and constant tweaking is needed, but if you force it to stick to well-defined design patterns, the models can produce high-quality, easily maintainable code. He reviews every line, and the code is as good as he’s ever created, he says: “I’ve found it absolutely revolutionary. It’s also frustrating, also hard, a different way of thinking, and we’re only getting used to it.”
