The Half Life of Agent Infrastructure — Ben Kus, Box
Agent architectures are expiring quickly. Keep model, search, and orchestration choices replaceable, and evaluate platforms by how well they handle repeated change.
Box operates at **over an exabyte of data**, tens of millions of users, and roughly **a trillion tokens**. Its CTO says last year’s graph-based agent pattern was overtaken by looping agents, while adaptive model selection and agentic search are now preferred approaches.
Treat agent infrastructure as replaceable. Separate components, make migrations routine, prepare teams for repeated rewrites, and choose vendors partly by how they handled transitions over the previous six to twelve months.
Box operates at **over an exabyte of data**, tens of millions of users, and roughly **a trillion tokens**. Its CTO says last year’s graph-based agent pattern was overtaken by looping agents, while adaptive model selection and agentic search are now preferred approaches. Treat agent infrastructure as replaceable. Separate components, make migrations routine, prepare teams for repeated rewrites, and choose vendors partly by how they handled transitions over the previous six to twelve months. This advice comes from enterprise-scale experience, and today’s preferred stack is not presented as durable. Frequent rebuilding also creates breakage, skepticism, and morale costs that modular architecture alone cannot remove.
This elevates replaceability from a model-level concern to the governing principle for the entire agent stack. It confirms that routing and harness policies are temporary, workload-dependent choices, while adding an organizational cost the technical candidates largely omit: recurring migrations can damage stability, trust, and morale even when components are modular.