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

Enterprise AI adoption is likely to move more slowly than Silicon Valley expects because security, operations, and economics remain unresolved. But the speakers warn that companies may be underestimating the opportunity by at least an order of magnitude. If agents eventually outnumber people by 100 or 1,000 to one, software will need to serve them as well as human users. That does not mean software companies should simply become APIs: agents can work through APIs, CLIs, MCP, or existing computer interfaces, and may choose different methods for different tasks.

A promising workflow gives coding agents access to SaaS tools, business processes, and relevant context so they can complete work through code or APIs. The speakers point to cloud co-work, OpenAI’s proposed super app, and Perplexity Computer as examples of this direction, where a single agent might coordinate work now split among many small tasks and a skilled person. The shift depends on people being able to direct agents, however, and many workers may find it difficult to describe their jobs as algorithms or flow charts. Tools can move workers up an abstraction level, but that has historically required new skills; managers learning spreadsheets illustrates how tools can also enable more iterations without requiring every task to be delegated.

Agents may combine existing tools with code they generate for unusual tasks. In an agent under development, the system can select an existing skill or Box tool, or write code on the fly; most work is expected to use established APIs, with generated code handling unforeseen document operations. Aaron Levie says Box rolled out an official CLI that can be provided to Claude Code for natural-language interaction with Box. The notes say Opus 4.6 can orchestrate tasks such as uploading a desktop folder or processing its documents. Yet shared repositories create coordination hazards: an agent could move or delete files while another user is writing to them.

Agents could also connect information across systems that IT teams have not integrated in advance, which may help with complex enterprise work spanning dozens of systems. But user-created integrations could damage systems of record, so read-only access may be a safer starting point. At a hypothetical company with 5,000 employees, agents could generate about 10,000 system interactions an hour, creating substantial demands for coordination and access management. Treating agents exactly like employees misses the need for agent-specific oversight and accountability: agents can expose information in their context and may be easier to social-engineer than people. Prompt injection could turn an agent into an effective information-leaking actor, making enterprises likely to restrict access to sensitive resources such as M&A data rooms even as startups move faster.

Enterprise software has persisted through technology shifts because it encodes organizational logic, not just software logic. Systems such as SAP embody domain knowledge in interfaces, middleware, and usage practices that vibe-coding alone may not reproduce. Still, agents’ need for broad data access could make repeated workflow friction a reason to replace systems that obstruct automation; software back ends may converge toward more generic APIs, while state, compatibility, policy, and security keep software layers necessary. Vendors will need strong APIs, agent identity and access controls, and viable monetization. Agent use might increase revenue for some products and reduce it for others, especially where revenue depends on users working inside an interface. Vendors’ use of agents on customer data is distinct from training models on that data. Agents that cannot access a vendor’s product could leave it nonviable, but paid recommendations could distort tool selection, and ad hoc agent-built systems could fragment data and create security and legal risks.

The speakers argue that AI’s financial potential may be understated when forecasts treat revenue as linear and markets as fixed-sum: easier access to PCs, cloud, and SaaS CRM expanded usage, and new agent-based services could do the same. One speaker says roughly 50 infrastructure companies in their portfolio with visibility had gone “asymptotic” in the previous six months. Meanwhile, token costs are a significant part of AI software’s cost of goods sold, and compute spending could materially affect earnings. Since research and development is roughly 14%–30% of revenue at public technology companies, CFOs may need to forecast engineering token budgets for Wall Street. Teams must balance long prompts, parallel work, speed, and experimentation against waste; one speaker favors tolerating substantial waste for now and says developers may use 90% of tokens trying multiple approaches.

More software and agents are expected to consume more computing resources, but the scale and timing are uncertain. Speakers disagree on whether the expense mainly repeats the move from capital expenditure to cloud operating expenditure or poses a new challenge because any end user can spin up resources. Data-center constraints currently limit tokens, and users of the Claude Code Max plan are reportedly being blocked after about three prompts. Capacity, algorithms, and hardware could ease the constraint; local compute or edge devices could also relieve cloud pressure, though when and how much is unclear. The speakers expect token limits and budgets to ease as capacity grows and compute gets cheaper, but say the next major efficiency breakthrough cannot yet be predicted.