AI and Proactive Reliability with Kolton Andrus

The discussion begins with the importance of reliability in complex systems, especially as AI begins to generate a significant portion of code. Kolton emphasizes that while AI can produce large amounts of code quickly, it lacks the 'scar tissue' — the intuition developed by engineers from experience with failures and outages. This gap could lead to more failures as code complexity increases, highlighting the need for new approaches to system reliability and proactive testing.

Kolton traces the origins of chaos engineering to his early work at Amazon, where proactive failure testing was developed to prevent outages. The concept was advanced at Netflix with the creation of Chaos Monkey, which was designed to intentionally induce failures in production environments to test system resilience. Over time, chaos engineering has evolved from basic 'pull-the-plug' experiments to sophisticated tools that allow safer, more comprehensive testing of dependent services and failure modes in cloud environments.

Initially, chaos engineering was practiced mainly by industry leaders like Amazon and Netflix. As cloud computing and microservice architectures became widespread, other companies started recognizing reliability as a systemic issue, prompting broader adoption. Kolton discusses the cultural challenges, such as rewarding firefighting over preventive work, and argues that moving toward proactive reliability requires changing incentive structures and organizational mindset—seeing reliability as a day-to-day responsibility rather than only reacting to failures.

AI accelerates the pace of software development, generating code and tests rapidly. However, Kolton warns that increased code volume leads to increased potential for failures if not accompanied by rigorous testing and validation. He stresses the importance of explicit safety, reliability guardrails, and validation within the AI development lifecycle, emphasizing that AI should support, not replace, human judgment in ensuring system robustness.

Bolton introduces the concept of 'reliability yardrails'—protocols to ensure AI-generated code and systems meet safety standards. He advocates for explicit testing, validation of failure modes, and safety mechanisms like automated revertibility. Building AI that understands domain-specific nuances and can test its own failures will be key, but human oversight remains essential to prevent catastrophic errors, especially at critical infrastructure layers.

The conversation touches on whether industry standards or regulations might be necessary to ensure safety and reliability as reliance on AI grows. Kolton suggests that accountability and organizational incentives are crucial, and parallels are drawn to security regulations. He notes recent outages and emphasizes the need for defense-in-depth, layered safeguards, and shared responsibility across organizations to mitigate the risks of faster, AI-driven development.

Kolton discusses the long-term prospects of AI in system reliability, acknowledging ongoing growing pains and the importance of cautious progress. He compares current challenges to early autonomous vehicle testing, emphasizing the need for rigorous checks before full deployment. The discussion concludes with a call for responsible innovation—balancing speed with safety, regulation, and a shared understanding of AI’s role in future engineering.