Every Agent Needs a Box — Aaron Levie, Box

What It Covers

Box CEO Aaron Levie joins Latent Space with Chroma CEO Jeff Huber to examine why enterprise AI agent deployment lags behind coding agents, covering data governance, agent identity management, access control architecture, context engineering challenges, and why Fortune 500 companies face a multi-year transformation timeline before realizing compounding productivity returns from autonomous agents.

Key Questions Answered

• •Agent Identity Architecture: Treating agents as standard user accounts creates critical security gaps. Unlike human employees, agents carry no legal liability, deserve no privacy protections, and require full auditability by their creator. Enterprises need a distinct identity layer — separate from Okta-style human IAM — that grants agents scoped file-system access, maintains creator oversight, and prevents unauthorized data exposure across organizational boundaries.
• •Coding Agent Advantage vs. Enterprise Gap: AI coding agents succeeded because of eight compounding advantages: full codebase access for new engineers, text-in/text-out medium, heavily trained models, developer self-use feedback loops, technical user base, and open knowledge sharing. Every other enterprise knowledge workflow — legal, finance, banking — faces six to seven structural headwinds against each of those properties, creating a multi-year deployment gap.
• •Context Engineering at Scale: A knowledge worker may have 10 million documents across teams and projects — roughly 50 million pages — but reliable model performance degrades significantly beyond approximately 60,000 tokens. Bridging that 50-million-to-60,000-token ratio requires purpose-built agentic search systems, multi-pass retrieval with self-ranking, and models capable of recognizing when continued searching will not yield better results rather than returning incomplete answers.
• •Workflow Adaptation Runs One Direction: Enterprises should not expect agents to conform to existing workflows. The coding world demonstrated that humans restructure their work to make agents effective — not the reverse. Organizations that proactively re-engineer documentation practices, digitize tacit knowledge, and restructure data access for agent readability will gain compounding velocity advantages over competitors still waiting for a frictionless drop-in solution.
• •Agent Evals as Core Infrastructure: Every enterprise deploying agents needs a private, held-out evaluation benchmark tied to their specific workflows — equivalent to Box's internal eval suite covering industries like financial services, legal, healthcare, and public sector. Running models against these benchmarks at each update cycle catches regressions, guides model selection, and validates harness changes. Box observed roughly 15-point score jumps between consecutive Anthropic Sonnet model generations on their internal suite.
• •Context Pruning Over Retention: Frontier models performing agentic search repeat failed strategies when unsuccessful attempts remain in the context window — even when the model's own reasoning trace flagged those attempts as flawed. The practical fix is active context pruning: remove failed search branches from the window entirely, but inject a brief summary noting the failure so the model avoids repeating it, rather than leaving the full error trace to re-anchor behavior.