In modern software development shops, the sound of mechanical keyboards has been replaced with the silent exchange of tokens.
This is the world of vibe coding, a term that Cameron Etezadi, CTO of LaunchDarkly, views with a mix of skepticism and wonder. After 30 years in the industry, Etezadi has seen the cost of writing code drop effectively to zero, yet he warns that this efficiency is a distraction from a much deeper, structural problem in software development.
“Vibe coding represents a fundamental shift in how we build,” Etezadi said in a recent SD Times “What the Dev?” podcast. “In the old world, code was deterministic. If you gave a computer a command, it executed that exact logic every single time, barring a hardware failure. Today, we work with probabilistic agents. Asking an AI the same question two days in a row might yield two different blocks of code.”
So AI, which has massively reduced the routine toil of development, has shattered the certainty of outcomes that businesses rely on. Organizations no longer need armies of specialists to ship a product; a single developer empowered by a fleet of agents can produce 10 times the output of a traditional team.
Organizations want to ship faster and more securely, and allow people to have confidence in the product being shipped, Etezadi said. Feature flags, Launch Darkly’s original product for feature-level control, has to be built upon with a feedback loop that allows you to release, observeand iterate, and that feedback loop includes things like experimentation. “I always use the case of a checkout pipeline, because early in my career, I owned Amazon’s entire checkout process, which is a very weighty responsibility,” he said. “Does the blue ‘Buy Now’ button convert better than the yellow button That’s the experiment. “
The speed with which code is developed, however, creates an issue of trust. A recent study by LaunchDarkly revealed that while 94% of companies are shipping code faster than ever, 91% are less confident in what is actually going out the door. “The problem is that while AI produces code that is structurally correct —- it compiles and looks right -— it is not always behaviorally correct,” Etezadi said. He noted that in large-scale systems, a slight optimization choice made by an AI could lead to disastrous performance issues, such as an algorithm running in polynomial time that slows to a crawl as the user input got bigger and bigger. “It was bad coding, but it looks structurally correct.”
The solution, he said, isn’t to fight the “vibe” but to govern it. “Since you cannot fire an AI for a mistake, you must surround it with a runtime control plane. This involves using “guardrails” like feature flagging, observability, and experimentation,” Etezadi said. By treating AI-generated code as a probabilistic variable, companies can use tools like “AI configs” to nudge outcomes back toward determinism. If an agent starts drifting into inefficient territory, the system must be able to detect the latency in real-time and kill the feature instantly.
Ultimately, he said, vibe coding requires shipping smaller, more frequent updates while automating the validation processes that humans can no longer keep up with. Today, he said, “I’ve never run on the hamster wheel this fast in my entire career.”
To listen to the entire conversation, click here to listen now.
