The AI Bottleneck: Why Speed Without Control is a Liability

Cameron Etezadi, CTO at LaunchDarkly, argues that the rapid adoption of AI-generated code has changed where the real bottleneck lies in software delivery. While writing code is now nearly free, the downstream tasks of running, monitoring, and taking responsibility for that code have become more difficult. Etezadi calls this "vibe coding," where developers produce code that runs but ignores the realities of production. For engineering leaders, the competitive edge no longer comes from raw speed, but from building control systems that manage the unpredictability of AI. Those who transition from individual contributors to managers of AI agents will outpace their competitors by turning operational friction into a long-term advantage.

The Hidden Cost of Vibe Coding

The industry is currently focused on the ease of generating functional code. However, Etezadi notes that AI models often miss second-order effects, such as the architectural decisions that determine if a system survives in production. An AI might generate a solution that passes unit tests in a controlled environment, but it often ignores variables like configuration or scaling limits.

"Vibe coding kind of takes us, takes that to the next level, to be honest. It is very easy to produce code that works, especially down a happy path where you achieved your objective as long as everything is perfect in the way. It does not scale, it does not operate, it does not run at enterprise scale."

-- Cameron Etezadi

The result is a shift in the bottleneck: the time saved writing code is spent debugging systems that behave incorrectly. Etezadi points out that the best engineers are no longer just writers of code. They are thinking partners who use AI to play adversarial roles, questioning assumptions and finding architectural flaws that only appear under production load.

The New Engineering Archetype: From IC to Frontline Manager

The traditional two-pizza team is becoming outdated. Etezadi suggests that the modern engineering unit is shrinking, with a single engineer acting as a manager for multiple AI agents. This structure requires a shift in mindset: the engineer must handle work breakdown, project management, and context switching at a level previously reserved for team leads.

The risk is that AI lacks self-reflection and accountability. When a system drifts, the AI does not lose sleep, but the human operator does. Etezadi emphasizes that as regulations and corporate liability increase, the responsibility for code remains with the human. The competitive advantage goes to those who treat AI as a tool for experimentation rather than a replacement for engineering judgment.

"I think the behavior of an engineer has changed from I am an IC to I am a frontline manager. And so you have to behave on project management on work breakdown like a frontline manager to be successful now."

-- Cameron Etezadi

Where Immediate Pain Creates Lasting Moats

Most teams view speed as the primary metric of success, but Etezadi warns that speed without a control plane is a disaster waiting to happen. In a world where AI introduces randomness, the ability to ship small, reversible changes becomes the defining trait of a high-performing organization.

This requires investing in lane bumpers: feature flags, monitoring, and automated experiment loops. While these controls introduce friction that many teams avoid, they create a durable advantage. By building systems that detect drift and allow for immediate course correction, teams avoid the stress of production failures and move faster than those who prioritize raw output over stability.

Key Action Items

• Audit Your Hiring Pipeline (Immediate): Shift interview focus from algorithmic memorization to debugging and code review. Require candidates to use AI to generate solutions, then hold them accountable for explaining the code and identifying its flaws.
• Implement AI Adversaries (Next 30 Days): Move beyond using AI as a one-shot coder. Build workflows where one model critiques the output of another, specifically looking for scaling bottlenecks and security vulnerabilities.
• Adopt the Two-Slice Team Mindset (Next Quarter): Evaluate your team structure. Shift from large, collaborative squads to smaller, high-autonomy units where the lead acts as a manager of agentic workflows rather than a traditional individual contributor.
• Formalize the Observe-Iterate Loop (Next 6-12 Months): If you are not using feature flags to manage releases, you are flying blind. Invest in the control plane infrastructure that allows you to treat every deployment as a reversible experiment.
• Prioritize Meatball Metrics (Ongoing): Stop optimizing solely for velocity. Track Mean Time Between Loss of Sleep (MTBLOS). If your team is constantly waking up to fix production issues, your speed is a liability, not an asset.