Context engineering is the critical enterprise bottleneck — 60K usable tokens against millions of pages¶
Insight: According to Aaron Levie (Box CEO), the critical bottleneck for enterprise AI agents is context engineering: with approximately 60,000 usable tokens against 50 million pages of organizational data, search quality and ranking become make-or-break challenges. Agents require distinct identity profiles from employees (creator liability, not privacy protections) and organizations must fundamentally restructure documentation, data organization, and access controls.
Detail: Levie explains why coding agents succeeded first: broad codebase access, text-in/text-out workflows, developer technical literacy, and tight feedback loops with AI lab creators — advantages that don't transfer to enterprise knowledge work. Organizations must "adapt to how the agent works" rather than expecting agents to adapt to existing processes. The "knowledge capture premium": companies documenting institutional knowledge gain compounding advantages in both agent effectiveness and employee onboarding.
Sources
Related: context-engineering-supersedes-prompt-engineering