Descriptions:
Box CEO Aaron Levie joins Latent Space—guest-hosted by Chroma CEO Jeff Uber—to make the case that enterprise content management is becoming critical infrastructure for the agentic AI era. Levie’s argument: the 20+ years of corporate files stored in platforms like Box (contracts, research materials, memos, product roadmaps) have historically sat dormant between active human use, but AI agents transform that passive archive into a continuously valuable knowledge layer—one that new employees, sales teams, autonomous agents, and background workflows all need to access simultaneously.
A significant portion of the conversation covers Box’s Apex agent evaluation framework, which tests AI models on complex question-answering tasks across industry-specific document sets representing typical enterprise workspaces (legal, investment banking, and similar verticals). Levie reports 15-point benchmark jumps between successive model generations—citing Claude Sonnet 4.6 versus Sonnet 4.5 as a specific example—as evidence that model capability improvements are now directly measurable in enterprise document tasks. The episode also introduces context window management as an emerging engineering discipline: Levie and Uber discuss research showing that leaving errors in an agent’s context causes the model to repeat mistakes even when it knows the approach failed, making active context pruning a necessary practice for long-running agents.
The discussion of OpenClaw’s acquisition by OpenAI frames the broader trend: the industry is converging on sandboxed, permission-scoped agent environments, and Levie argues every such agent will need a governed data layer—which is exactly what Box is positioning itself to provide.
📺 Source: Latent Space · Published March 05, 2026
🏷️ Format: Interview
