The Agent Era: Building Software Beyond Chat with Box CEO Aaron Levie

In this episode of The a16z Show, host and A16Z general partner Martin Casado is joined by Box CEO Aaron Levie and A16Z board partner Steve Sanofsky to explore the transformative impact of AI agents on enterprise software. The conversation centers on the emerging 'Agent Era,' where software must be reimagined not for human users but for autonomous agents that could outnumber humans 1,000 to one. The panel debates whether agents will flatten existing systems or instead reinforce and evolve them, emphasizing that agents are not interested in interface polish but in backend reliability, cost, and durability. They discuss the growing tension between enterprise caution and startup agility, with startups rapidly adopting agent-driven workflows while large organizations struggle with legacy systems and security concerns. A major theme is the economic disruption ahead: the cost of AI compute is poised to become the most critical budgeting challenge for engineering teams, with CFOs facing unprecedented pressure to allocate resources for tokens and infrastructure. The discussion also touches on the risk of agent-driven chaos—such as accidental data leaks, rogue integrations, and fragmented 'shadow systems'—and the need for new standards, access controls, and business models. Ultimately, the hosts argue that the real opportunity lies not in building for agents as a marketing tactic, but in building better, more agent-friendly systems that can scale with the new reality of AI-driven work. Key takeaways include: 1) Software must be built for agents, not just humans, with APIs and system reliability becoming the new competitive moat; 2) The economics of AI are misunderstood—companies are off by at least an order of magnitude in estimating the opportunity, and usage-based pricing will dominate; 3) The biggest challenge isn’t AI capability but managing the explosion of compute costs and agent-driven resource consumption; 4) Enterprises will face a 'cat and mouse' game with agents, requiring new governance models that treat agents as extensions of humans; 5) The future of software is not about eliminating layers but about evolving them to serve agents, with the most successful companies being those that enable seamless, secure, and scalable agent access to data and workflows.