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

Episode Summary

Executive Summary: The discussion argues that AI agents will reshape software through a slower, more layered diffusion than Silicon Valley expects. The speakers agree that enterprises will build for agents via APIs, CLIs, and computer use, but disagree on how much existing systems will change. They emphasize new economics, usage-based pricing, agent identity/security, compute budgets, and the likelihood that agents will amplify demand rather than collapse software layers.

Main Topics: AI diffusion will be slower than Silicon Valley expects (Priority: 5/5): The speakers argue that AI capability will spread unevenly because most enterprise knowledge is embedded in systems, workflows, and tacit domain expertise that cannot be replaced quickly by vibe coding or generic models. Agents as a new software consumer layer (Priority: 5/5): A central theme is that software must increasingly support agents as first-class users. Agents will interact through APIs, CLIs, MCP, and computer-use workflows, often acting as orchestrators of existing tools rather than replacing them outright. Layers persist; they rarely disappear (Priority: 5/5): The conversation repeatedly returns to historical precedent: new technology waves add abstraction layers rather than removing them. Agents may shift where work happens, but ERP, SaaS, and systems of record will remain because they encode organizational logic and compatibility constraints. Enterprise security, identity, and shared-state problems (Priority: 5/5): A major practical concern is how to govern autonomous agent activity in shared enterprise environments. The speakers debate permissions, oversight, liability, privacy, and whether agents should be treated like human users or as extensions of them. Economics, usage-based pricing, and compute budgets (Priority: 4/5): The group focuses on how AI changes cost structure. Token usage, engineering compute budgets, and usage-based monetization could reshape SaaS economics, but the speakers believe current models are still too small and linear. Agents may choose software based on backend quality, not UI polish (Priority: 4/5): One speaker argues that agents care more about durability, cost, reliability, and data/model compatibility than interface aesthetics. This shifts competitive advantage toward stronger infrastructure and better systems of record. New business models and market expansion (Priority: 4/5): The speakers suggest AI agents will unlock new forms of consumption, micropayments, and service businesses by enabling previously uneconomic transactions and generating much more software and data usage overall.

Key Arguments: AI adoption will take longer than many expect because domain knowledge is not fully captured in data layers or interfaces; enterprise software is too complex to replace quickly. Agents will increasingly need to use existing software rather than replace it, making APIs, CLIs, MCP, and computer use essential. The right abstraction is not 'marketing to agents' but building systems agents can reliably operate; agents pick backends based on semantics like cost and durability. Historical precedent suggests software layers persist and new ones are added, not collapsed; the same is likely true for agents. Human users are often the bottleneck in exploiting software capability, but agents can surface and use more of the stack's latent functionality. Enterprise adoption will be constrained by identity, access control, liability, and shared-state risks when humans and agents collaborate in the same systems. The AI transition will likely expand total software and compute consumption, making current revenue and token models too small and too linear. Usage-based pricing and token costs will force companies to rethink engineering budgets, finance controls, and product monetization. Some startups and new services businesses can be built first-principles for agents, but incumbents like SAP and Workday will remain important because their domain logic and workflows are deeply embedded. In the long run, agents may pressure software vendors to improve APIs, interoperability, and machine-readable access, because agent effectiveness will drive business performance.

Data Points: Agent-to-employee ratio: 100x to 1000x more agents than people - Used repeatedly as the hypothetical scale at which software must be built for agents. Marketing team example: 50 marketing people - Illustrative example used to show that only one person may understand the full workflow well enough to document it. Deprecated/current tool usage example: 7 apps on an iPhone - Used to argue that humans dislike learning too many interfaces, while agents may not have that constraint. Enterprise team example: 5,000 employees - Used in a hypothetical shared-repository scenario to illustrate scale and coordination problems with agent activity. Agent activity example: 10,000 times an hour - Describes possible runtime interaction volume against shared systems in a large company. Cohort example: 2 years - Used in the spreadsheet analogy to show how a new abstraction layer eventually becomes mainstream within a job family. Iteration example: 30 iterations - Contrasted with older workflows where people could only do a couple of iterations before shipping. Demand example: 5 to 10 people - Anthropic growth marketer example claimed one person could automate work that might otherwise require several people. Portfolio visibility: 240 companies - One speaker says they invest in roughly this many companies and have visibility into about 50. Time window: last 6 months - The speaker says infrastructure companies in the portfolio have gone 'asymptotic' over this period due to increased software consumption. R&D spend range: 14% to 30% of revenue - Referenced as the typical range for public technology companies when discussing how compute costs may affect EPS. Cloud business example: 60,000 units a year - Used to describe the old server market before cloud expansion. AI budget uncertainty: 1% to 100% - Used rhetorically to show the wide disagreement on how much engineering spend should go to tokens/compute. APIs/finops analogy: 2016 to 2018 - Referenced as a prior period when FinOps emerged to manage cloud/API spend.

Pivotal Quotes: "The diffusion of AI capability is going to take longer than people in Silicon Valley realize." — Narrator / opening framing: Sets the thesis that AI adoption will be slower and more constrained by enterprise reality than hype suggests. "If you have a hundred or a thousand times more agents than people, then your software has to be built for agents." — Aaron Levy / Box CEO: Core argument for why enterprise software interfaces and architecture must adapt to agent-first usage. "I actually think that's almost exactly wrong." — Martin Casado / A16Z GP: Pushback against the idea that software should merely be marketed to agents; he argues agents will select better backend systems based on semantics.

Implications: Expect slower but deeper AI adoption: more agent-driven usage, more compute consumption, stricter identity/security controls, and pressure on vendors to expose better APIs and machine-readable workflows. The winners will be software systems that agents can reliably operate.