Agentic infra changes everything (Interview)

What It Covers

Adam Jacob discusses how agentic AI systems have fundamentally changed infrastructure development at System Initiative, leading him to delete five years of UI work. He covers the AWS outage, AI bubble economics, and why agents are glue not magic.

Key Questions Answered

• •Infrastructure Repatriation Economics: AI workloads require GPU-dense data centers with bare metal compute and fast networks, creating gravity that pulls other workloads on-premises. This reverses cloud migration trends as companies realize they can run traditional workloads more efficiently on infrastructure they already need for AI, changing deployment economics fundamentally.
• •Agent-Driven Development Loop: Building with AI agents enables parallel workflows where developers initiate multiple tasks simultaneously rather than sequential coding. The agent SDK from Anthropic wraps the entire control loop into a simple query API, eliminating need to manage turns, memory, or orchestration—developers just define tools and await results.
• •LLM Hallucination Correction: System Initiative corrects AI hallucinations immediately during code generation by using strict modeling language that validates properties in real-time, rather than waiting for compile or lint stages. This tight feedback loop dramatically improves output quality by catching errors at injection point, not post-generation.
• •Interface Design for Agents: APIs optimized for LLMs differ fundamentally from human-facing APIs—agents perform better with wide, exploratory interfaces rather than narrow, specific endpoints. One-to-one mappings of existing systems to AI tools fail; successful implementations expose AWS resources exactly as AWS describes them, leveraging training data.
• •Agent Autonomy Principles: Agents must earn autonomy through repeated successful performance under human observation before operating independently. The practical application uses change sets and policy engines to prevent YOLO infrastructure changes, requiring human review loops until systems prove reliable enough to reduce oversight gradually over time.