Summary
Ben Kus, CTO of Box, delivered this talk at the AI Engineer conference exploring what he calls the “half-life” of agent infrastructure — the idea that the best architectural choices for building AI agents are changing faster than any prior technology wave. Speaking from Box’s vantage point of over an exabyte of stored data, tens of millions of users, and approaching 10 trillion tokens of AI usage, Kus argues that the conventional enterprise wisdom of picking a solid technology stack and optimizing it over time is now actively dangerous advice.
Kus traces the rapid succession of leading approaches across two critical dimensions: agent orchestration (from graph-based traversal frameworks to bring-your-own harness patterns) and retrieval (from BM25 keyword search through embedding-based approximate nearest neighbor, to graph retrieval, to hybrid lexical-semantic search with rank fusion, and now to agentic search where models navigate data using their own intelligence). He identifies the Opus 4.0-to-4.5 transition as a meaningful inflection point that made robust instruction-following at scale practical and effectively opened the current era of production agent deployments.
The central takeaway is a call for architectural humility: enterprises should design systems that absorb change rather than optimize for today’s best answer, because the trillion-dollar investment flowing into AI is compressing the half-life of every infrastructure decision. Kus contrasts this with other long-stable engineering domains — database design, identity access control — where advice from several years ago still holds. For AI agent infrastructure, he suggests, it does not.
📺 Source: AI Engineer · Published August 29, 2026
🏷️ Format: Deep Dive
