In modern software development shops, the sound of keyboards being used has been replaced by the silent exchange of tokens.
This is the world of vibe coding, a term Cameron Etezadi, CTO of LaunchDarkly, views with a mixture of skepticism and wonder. After 30 years in the industry, Etezadi has seen the cost of writing code effectively drop to zero, yet he cautions that this efficiency is a distraction from a deeper, structural problem in software development.
“Vibe coding represents a fundamental change in the way we build,” said Etezadi in the recent SD Times “What the Dev?” a podcast. “In the old world, code was the goal. If you gave a computer a command, it always made sense, barring hardware failure. Today, we work with probabilistic agents. Asking an AI the same question two days in a row might produce two different blocks of code.”
So AI, which has greatly reduced the hard work of development, has disrupted the certainty of results that businesses rely on. Organizations no longer need armies of professionals to ship product; a single engineer empowered by a network of agents can produce 10 times the effect of a traditional team.
Organizations want to ship quickly and safely, and allow people to have confidence in the product being shipped, Etezadi said. The included flags, Run Darkly’s first product for feature level control, should be built on a feedback loop that allows you to release, view and iterate, and that feedback loop includes things like testing. “I always use the exit pipeline, because at the beginning of my career, I was in charge of the entire payment process for Amazon, which is a very difficult responsibility,” he said. “Does the blue ‘Buy Now’ button convert better than the yellow button That’s the test.
The speed at which the code is being developed, however, creates a problem of trust. A recent study by LaunchDarkly revealed that while 94% of companies are shipping code faster than ever, 91% are less confident about what actually gets out the door. “The problem is that while AI generates code that’s structurally correct — it compiles and looks correct — it’s not always behaviorally correct,” Etezadi said. He noted that on large systems, small optimization choices made by AI can lead to catastrophic performance problems, such as a polynomial-time algorithm that slows down as user input gets larger. “It wasn’t good coding, but it looks structurally correct.”
The solution, he said, is not to fight “chaos” but to rule. “Since you can’t accidentally shoot an AI, you have to surround it with a runtime control plane. This includes using “guardrails” like flagging, detection, and inspection,” Etezadi said. By treating AI-generated code as a possible variable, companies can use tools like “AI configs” to push results back into decision making. If an agent starts to drift into a malfunctioning area, the system must be able to detect the delay in real time and kill the defect immediately.
Ultimately, he said, coding the vibe requires deploying smaller, more frequent updates while automating validation processes that people can no longer keep up with. Today, he said, “I’ve never run a hamster wheel so fast in my entire career.
To listen to the entire interview, click here to listen now.
