What Breaks When Code Is Free?

We sat down with Cameron Etezadi, CTO at LaunchDarkly, who has spent 30 years building and leading software organizations at companies including Microsoft, Amazon, Google, and HashiCorp. We discussed what happens when AI makes code abundant, where the new engineering bottlenecks emerge, and how LaunchDarkly is building the control layer for safely releasing and managing AI systems in production.

The supply of code has changed faster than most of the systems around it. And “taste” is the new moat. Given that LaunchDarkly is the software that allows you to control at a granular level what is actually shipped and put in front of customers, you can think of it as a way to control taste at scale.

Which changes should reach customers? How should they be rolled out? Who validates the work? How does a company know when a release is causing harm? And when an incident happens, how quickly can the team contain it?

LaunchDarkly is experiencing this shift inside its own engineering organization. Cameron says the team is writing code roughly three times faster and shipped more than five million lines in six months. A legacy targeting UI rewrite that might once have required eight engineers and two years was completed by two senior engineers in three months with AI. The project was about 12 times cheaper by Cameron’s estimate, but the larger benefit was gaining 21 months in the market.

That example shows both sides of the code glut: AI can unlock projects that would otherwise remain stalled for years… but it also produces more pull requests, larger changes, and more output than human reviewers can comfortably process.

The problem becomes even more difficult when the software itself contains AI agents. Traditional software generally changes when a team deploys new code, but an agent can behave differently because its model was updated, its prompt drifted, its context or source data changed, or its probabilistic output took a different path. In other words, a release that worked during testing may produce a different result tomorrow without a conventional deployment.

This is where LaunchDarkly is uniquely positioned. The company was built around separating deployment from release and giving engineering teams control over live software behavior. Its feature-management infrastructure now performs 60 trillion evaluations per day and can deliver changes globally in under 200 milliseconds. That existing control layer provides the foundation for managing a world in which both the volume of code and the variability of software behavior have increased.

LaunchDarkly is extending that foundation through AgentControl. Teams can introduce agent changes gradually, target specific groups, compare prompts and models, reconstruct which combination of model, prompt, context, and policy produced a decision, and shut down harmful behavior without waiting for a new deployment.

Observability is part of the answer, but a dashboard alone is not control. A company needs to know that an incident is happening, understand what caused it, and have a direct path from that signal to an intervention. If customers experience the failure before a person notices the graph, the response loop is already too slow. LaunchDarkly’s advantage is its ability to connect detection with action at the speed production AI requires.

The Rest of the Discussion

• Dogfooding AI before building AI products. Cameron explains why he prefers “eating your own dog food” to “drinking your own champagne.” LaunchDarkly had to become skilled at using AI internally before it could build useful controls for customers. Teams use Claude Code, Devin, Cursor, OpenAI models, and other tools while gradually learning which option fits each job.
• Giving teams freedom while controlling AI spending. Rather than restricting an engineer because their token usage crossed a threshold, Cameron uses unusual spending as a reason to investigate. A $1,000 day might reveal waste, but it might also mean someone is solving one of the company’s hardest and most valuable problems. His goal is to corral spending and share what teams learn without forcing every employee into one tool.
• Measuring AI ROI through time gained. The legacy targeting UI rewrite was roughly 12 times cheaper by Cameron’s estimate, but he sees the 21 months gained as the more important result. Completing the work earlier gives LaunchDarkly more time to improve the product and create value in the market.
• Using the right model for the right job. Cameron compares open-weight models with Linux: they do not need to replace every proprietary product to become foundational. Linux did fantastically well as a software, but Microsoft still exists. Lower-cost models can perform evaluations, answer questions over a defined corpus, call tools, and handle routine analysis, while high-end specialized models remain valuable for difficult engineering work.
• Budget-aware model routing. A company can define an AI budget and automatically move traffic to a less expensive model when usage reaches a threshold. This allows teams to manage cost while preserving service, even if they accept a temporary reduction in quality.
• Why goal-seeking agents need business-aware guardrails. An agent measured only on customer satisfaction may learn that giving every customer a full discount produces excellent scores. Here’s your paperclip, sir. Cameron uses examples like this to show why goals, permissions, context, and intervention rules must reflect the whole business outcome rather than a single metric.
• The pull request becomes the next bottleneck. AI is producing more code, larger changes, and more pull requests than human reviewers can comfortably process. Cameron expects software development to become a series of agent-managed steps with humans brought in when risk, complexity, judgment, or taste make their involvement valuable.
• Determining when a human belongs in the loop. Cameron calls this the “secret sauce.” Some changes are easy to observe and reverse, while others require human judgment. Future systems will learn from code complexity, prior bugs, business priorities, and a company’s risk history to decide where review is necessary.
• Keeping sales and marketing informed when engineering ships constantly. Faster production creates an organizational signal-to-noise problem. LaunchDarkly addresses it through weekly team demos and AI-generated summaries that turn technical updates into concise material sales and marketing teams can use with customers.
• From developer to systems thinker. Cameron believes the builder’s identity is shifting from “I write code” to “I solve customer problems.” Product thinking, design, system interactions, judgment, and taste become more important as implementation takes less time.
• From theoretician to experimentalist. When teams can produce five versions in a day, they no longer need to predict the perfect solution before building. They can release controlled experiments, define what they want to learn, measure outcomes, and use evidence to guide the next iteration. Cameron describes this as a form of “ultimate agile.”

Learn more about LaunchDarkly and follow Cameron Etezadi on LinkedIn for more of his perspective on AI-native software development. You can also watch the AgentControl demo to see how LaunchDarkly helps teams evaluate, release, and control AI agents in production.

‍