Agent-Driven Enterprise Software Requires Durable, Cost-Efficient Systems

The Agent Era: Beyond Chatbots and Towards a New Enterprise Paradigm

This conversation with Aaron Levie, CEO of Box, reveals a fundamental shift in how enterprise software will be built and consumed: the rise of agents as primary users. The non-obvious implication is not simply a new interface, but a complete re-architecting of systems for durability, cost-efficiency, and reliability, moving beyond human-centric design. Those who grasp this will gain a significant advantage by building for the future, anticipating the needs of a thousand-fold increase in automated users. This is essential reading for anyone involved in software development, product strategy, or enterprise IT who wants to understand the seismic shifts underway and position themselves ahead of the curve.

The Agent's Choice: Prioritizing Substance Over Style

The prevailing wisdom often suggests that technological advancements, particularly in AI, will lead to simplification and the elimination of intermediaries. We've seen this narrative play out with marketplaces and SaaS, where the promise was a flatter, more efficient system. However, the arrival of agents is proving this assumption flawed. As Levie points out, agents do not seek simpler systems; they demand better ones. Their selection of back-end tools is driven by pragmatic considerations like durability, cost, and reliability, not by the polish of a user interface. This is a critical divergence from human interaction, where aesthetics and ease of use often hold sway.

"Agents do not want simpler systems they want better ones they choose back ends based on durability cost parameters and reliability not interface polish"

This shift has profound implications for software companies. The question is no longer whether to support agents, but what it means when agents vastly outnumber human employees. The "Agent Era" necessitates a fundamental re-evaluation of product design. Instead of focusing solely on human interfaces, companies must now consider agent interfaces -- APIs, CLIs, and other programmatic access points. The success of "co-working" agents, which can understand data and then code their way through tasks using APIs, exemplifies this paradigm. This capability moves beyond mere data consumption to active task execution, a superpower that Levie believes is beginning to compound.

The Shifting Tides of Abstraction: From Specialists to Coordinators

The history of computing is a story of abstraction layers, each building upon the last. Levie draws a compelling parallel to the introduction of spreadsheets. Initially, complex financial modeling was a painstaking manual process. With spreadsheets, the task became more accessible, and a new cohort of "spreadsheet people" emerged. These individuals didn't necessarily become spreadsheet developers but rather skilled users who could leverage the tool to coordinate and execute tasks previously done by many. The job moved up a rung.

This pattern is repeating with agents. The "entropic growth marketer" tweet, illustrating one person automating the work of five to ten siloed roles using AI tools, highlights this. This individual wasn't just a user; they were a systems thinker, capable of orchestrating multiple agents to achieve a complex outcome. This suggests that the future workforce will increasingly resemble this marketer -- individuals who can effectively coordinate and direct AI agents. The "rocket science" part of setting up and managing these agents is rapidly diminishing, leaving behind a need for domain expertise and the ability to articulate complex workflows.

"The rocket science part of it just is going to evaporate in very short order and then you're talking about wow there's a giant chunk of domain expertise that they don't"

The distinction between using an agent as a "computer" versus a "coder" is also crucial. While some agents can generate code on the fly, many are proving more effective at using existing tools and APIs as their "computer." This means the focus for software companies shifts from building AI-generated code to ensuring their systems are robust, accessible, and well-suited for agent interaction. The development of a Box CLI, allowing natural language interaction with the entire Box system, exemplifies this approach. It empowers agents (and humans) to orchestrate complex operations like uploading and processing entire folders, a task that would have previously required significant manual effort or custom scripting.

The Enterprise Paradox: Control vs. Agility in the Agent Economy

The proliferation of agents introduces a new set of challenges for enterprise security and control. Unlike human employees, agents offer complete oversight. This leads to a paradox: how do you grant agents the autonomy to be effective without creating unmanageable risks? The traditional approach of treating agents like humans, with their own identities and permissions, quickly breaks down. Levie argues that agents are fundamentally an extension of the user, and the enterprise must retain this oversight.

"The default end to end argument is you treat them like human beings and you use it doesn't work so you can't fully treat them like humans because here's the thing and with regular humans you don't get to look at the slack channel of the person that that is working with you or working for you you don't get to log in as them you don't get to oversee them"

This need for oversight clashes with the very nature of agent autonomy. The risk of prompt injection, where an agent can be tricked into revealing sensitive information, is a significant concern. This has led to a cautious approach within enterprises, with many opting to "close everything off" until a greater sense of sanity prevails. However, this creates a stark divide between large, cautious enterprises and agile startups. Startups, unburdened by legacy systems and security concerns, can move much faster, leveraging agents to innovate and outpace their larger counterparts. This dynamic is likely to lead to a period where individuals and startups, particularly developers, will have a significant advantage over established enterprises.

The Data Dilemma: Systems of Record in an Agent-First World

A core tension emerges around "systems of record." Traditionally, enterprise software vendors sold not just data, but the intelligence, domain expertise, and systems built around that data. Now, agents primarily want access to the data itself, and they want it licensed with unlimited access. This challenges the long-standing business models of companies like Workday and SAP, which have historically controlled access to their data.

The fear is that agents, given free rein, will fragment data and create de facto new systems of record, bypassing established IT structures and potentially introducing security vulnerabilities. While some argue that agents will simply map to existing organizational and regulatory boundaries, the potential for disruption is significant. Startups built from the ground up with agents in mind, free from legacy constraints, are likely to prove highly disruptive. They can offer agents unfettered access to context and the ability to write software on the fly, creating new business models based on previously underutilized information and software.

Actionable Takeaways: Navigating the Agent Era

• Embrace Agent-Native Design: Prioritize building robust APIs and programmatic interfaces over solely focusing on human-centric UIs. This is not about marketing to agents, but building systems they will choose. (Immediate)
• Develop Systems Thinking Capabilities: Foster a culture where individuals can think about their work as a system that can be orchestrated by agents. This skill will become increasingly valuable. (Ongoing Investment)
• Re-evaluate Security Models: Move beyond traditional human-centric security. Develop strategies for agent oversight, data access control, and prompt injection mitigation. (Immediate Investment)
• Invest in Agent Orchestration Tools: As agents become more prevalent, tools that help manage, coordinate, and monitor them will be critical. (12-18 Months)
• Focus on Data Accessibility: For enterprise software vendors, shifting towards more open and accessible data models will be crucial for agent adoption. (This pays off in 18-24 months)
• Experiment with Agent Workflows: Encourage experimentation with agents for internal tasks, even if it means a controlled "waste" of tokens, to discover new efficiencies and capabilities. (Immediate)
• Prepare for Economic Shifts: Understand that the economics of enterprise software will change. Businesses that can offer agents efficient access to data and functionality will thrive. (Long-term Investment)