Aaron Levie
Co-founder and CEO of Box; thinks clearly about what it means to build software for agents rather than humans.
Last updated: 2026-04-13
Overview
Aaron Levie co-founded Box in 2005 and has led it through several platform transitions — from web file storage, to enterprise content management, to an AI-native document platform. His perspective on AI agents is shaped by running a company whose product is a natural fit for agent-centric workflows (“every agent loves working with files”) while also confronting enterprise-grade governance challenges at scale.
His core thesis for the current era: software companies that survive the agent transition are those that build for agent consumption — API quality, semantic richness, and agent identity management — not just human UX.
Key Ideas
Software must be built for agents
If there are 100–1000x more agents than people interacting with software, then the agent interface to your tool needs as much design attention as the human interface. Business performance will correlate with how well agents can access what they need. The companies that win are those that build genuinely better systems — not better agent-facing marketing.
See agent-first-software.
Semantic quality beats interface quality
Counter-intuitive: agents don’t choose tools based on documentation quality or clean API design. They select based on the actual technical properties of systems — cost, durability, semantics. The path to being chosen by agents is building better systems, not better IDLs or landing pages.
Agents as extensions of you, not employees
Agents are not independent workers — they’re extensions of the human who runs them. Full liability, full oversight, no privacy rights. The practical implication: you must be able to “log in as” your agent. But this creates a paradox for multi-agent collaboration: if you can log in as your agent, it can’t securely hold information given to it by another agent.
Keeping something secret in a context window is currently an unsolved problem. Anything in the context can potentially be prompt-injected out of it. Enterprise adoption is throttled until this is resolved.
The coding agent + SaaS access paradigm
The compounding pattern Levie sees: give a coding agent access to your SaaS tools and document workflows, and it becomes capable of not just reading/reasoning but of coding its way through arbitrary tasks using whatever API it needs. This “co-working” model (Box CLI + Claude Code being his concrete example) is the emerging paradigm.
The immediate engineering challenge this surfaces: coordination at agent scale. 5,000 employees with personal agents = potentially millions of operations per hour on shared systems. Concurrent writes, accidental deletes, permission boundary violations — problems that human-paced usage never triggered.
Diffusion gap
AI capability will diffuse to enterprises much more slowly than Silicon Valley expects. Legacy systems (SAP, ERP) encode domain knowledge in UI, middle tiers, and usage patterns — not just data. “It’s absurd to think you’re going to vibe code your way to SAP.” First-principles agent-native companies (knowledge-work services built from scratch) will emerge as case studies but won’t immediately replace incumbents.
Connections
- agent-first-software — his core framework
- agent-sandbox — per-user isolation is Box’s approach to the agent identity problem
- s3-first-architecture — file-centric storage as a natural agent substrate; “every agent loves working with files”
Sources
- The Era of AI Agents — video discussion, added 2026-04-13