Next-Generation Software Built for Trillion-Agent Scale
Original Article Title: Building for trillions of agents
Original Article Author: Aaron Levie, Box
Translation: Peggy, BlockBeats
Editor's Note: As large-scale model capabilities continue to advance, AI Agents are evolving from mere "chat tools" into digital labor capable of independently performing tasks. From coding and contract processing to financial auditing and research data analysis, Agents are beginning to permeate nearly every aspect of knowledge work.
With the number of Agents concurrently operating within enterprises far surpassing the employee count, the primary users of software may also shift from "human" to "machine." In light of this trend, software design, infrastructure, and even business models are undergoing transformation. This article takes "building software for Agents" as a focal point to discuss how the software form and infrastructure of the Agent era will evolve.
Note: The author of this article, Aaron Levie, is the co-founder and CEO of Box, an enterprise cloud storage company, and a long-time technology industry thought leader focusing on AI and enterprise software trends.
Below is the original article:
Over the past few months, a significant shift has been occurring in the Agent domain. Around the end of last year, we entered a phase where programming-based Agents were able to independently complete longer tasks without the need for frequent human guidance throughout the development process.
These Agents are no longer just chatbots equipped with simple tools. Today, they often have their own isolated computing environment where they can autonomously write and execute code to address encountered issues, directly access APIs and CLIs, interact with various systems, and possess their own file system and long-term memory capabilities, among others. With these foundational capabilities, coupled with the maturation of best practices around running Agents in an agentic harness and significant advancements in models for tool invocation and software development, we are beginning to see a possibility: Agents will be able to handle virtually any task assigned to them.
Initially, this architecture was primarily driven by a group of programming-based Agents such as Claude Code, Devin, Codex, Factory, Cursor, Replit, and more. However, recently, this pattern has transcended early tech circles and is starting to enter broader personal experience and knowledge work domains, such as Claude Cowork, Perplexity Computer, Manus, and, of course, OpenClaw. The latter has taken this direction even further—it can run 24/7 in a persistent environment.
As capabilities advance rapidly, Agents will be introduced into almost every field of work. They will be used to review every contract, handle a large volume of front-line customer support issues, audit corporate finances, comb through massive medical research to drive drug discovery, generate the majority of software code, create sales and consulting presentations, and even complete transactions on behalf of consumers on the internet. Overall, they will be involved in almost all economically valuable work in society.
Moreover, this is not just about doing what we are already doing today. Agents will enable us to do more, such as running complex simulations that were previously too costly, rapidly generating multiple prototypes for every idea because the cost of starting a project has significantly decreased, making it easy to stop a project; we will advance more projects simultaneously; we will also be able to analyze almost all data, no longer relying only on sampling.
Taking these trends together, it can be foreseen that in future organizations, almost every employee will have multiple Agents working for them. It is not hard to imagine an enterprise having 100 times or even 1000 times the number of Agents as employees. When tens of trillions of Agents are operating simultaneously, they will become the primary users of future software.
However, most software was originally designed for humans. This means that the form of software is likely to undergo a significant change. So what will happen next?
Creating Software "Agent-Worthy"
Paul Graham once summarized the principles of software entrepreneurship in an extremely simple phrase: Make something people want.
This idea has given rise to the most successful batch of software companies in the 21st century and has also driven a new product methodology — tools should be simple, user-friendly, easy to get started with, solve explicit problems, avoid obscure terms, and have clear pricing.
And now, this phrase may need to be rewritten as: Make software that Agents want to use.
Currently, the people who use Agents the most are often developers or users with strong technical skills, and they usually have their own tool preferences. But as Agents begin to handle various tasks for knowledge workers, this user preference will gradually weaken. Unless a unified tool is already specified within the organization, in many workflows, the real decision maker will be the Agent.
This means: they will decide which tools to use, what code to write, which libraries to call, and which skills to apply. Platforms that are more easily accessible to Agents and can better solve problems will gain an advantage more quickly than other products. Agents will not attend your online events or see your ads; they will only choose the most effective tools to complete tasks, and of course, you want that to be your product.
The key insight from this proposal is: everything must be API-centric (API-first).
If a feature doesn't have an API, it's almost as if it doesn't exist.
If a feature can't be invoked through CLI or MCP server calls, you are already behind.
If the API design is convoluted, with path conflicts that make it difficult for the Agent to use, then you are essentially willingly forfeiting the opportunity to be an Agent tool.
At Box, we are focused on building a file system oriented towards the Agent, so we are meticulously examining every detail of the API, considering where issues may arise in an Agent environment. This level of detail has historically only been seen in user experience (UX) design.
Just as designing software for users requires thinking from the user's perspective, designing software for Agents also requires the same approach. For example, Y Combinator's Jared Friedman has reminded developers: "Even the best developer tools mostly don't let you sign up via API. This was a huge problem in the Claude Code era because it meant Claude couldn't sign up for itself. Now, putting all account management functionality into the API should be table stakes."
If Agents can't easily sign up and start using your service, then in their eyes, you might as well not exist.
Changes in Business Models Will Follow
When Agents become the primary users of software, business models will also shift.
In some cases, models triggered by user seats that drive Agent execution may still apply. However, there will also be a significant number of Agents no longer tied to specific users, or their workload far exceeding traditional software usage. For instance, with just a few lines of input, an Agent could perform the equivalent of hours of human work within the software and only present the final result to the user.
As a result, the business models of some software products will evolve. Any tool looking to survive in the "Agent era" will need to introduce some form of usage-based or compute-based billing, even supporting Agents to handle payments on their own.
Next-Generation Infrastructure for Agents
Perplexity founder Aravind Srinivas once said, "Giving computers to humans was a good idea, but giving computers to computers to do work for us is a better idea."
As Agents have their own computing environment, can write and execute code, invoke skills to perform repetitive tasks, and integrate with various external tools and services, a whole new technical ecosystem will also emerge. Just as humans need on computers, Agents also need a similar but specifically designed infrastructure for them.
Some of these services will come from existing companies, as Agents still need access to existing data or need to collaborate between human users and Agents.
However, at the same time, a plethora of new product categories will emerge because the problems Agents face are entirely different from humans, and designing new services from scratch often makes more sense.
For example, Agents clearly need their own infrastructure, and the scale may be unprecedented. The next generation of hyperscale cloud platforms (or upgraded versions of existing giants) is likely to be built around this idea: the future data center will no longer primarily run our applications but will run our Agents instead. Companies like E2B, Daytona, Modal, Cloudflare, and others are already moving in this direction, and the scale of these sandbox computing environments may reach unprecedented levels.
Agents also need access to core enterprise files and manage their own data and memory to support long-running tasks. Similarly, enterprise systems also need to shift to API-first to allow Agents to access key data and services such as HR systems, CRM, workflow systems, data lakes, and more. Platforms that enable Agents to seamlessly operate on this data anytime, anywhere are most likely to take on future workloads.
Agents may also need their own identity system and the ability to communicate with each other. For example, Agent mail is providing email for Agents, giving them a persistent email address. Meanwhile, companies like Exa, Parallel, and others are also reimagining search engines to adapt to a world where "Agents are the primary search users." Many Agents also need to manage their own budgets, for example, by using wallets provided by Stripe or Coinbase for payments, which could even drive microtransactions to truly take off, allowing Agents to access paid tools and data resources.
Of course, security, compliance, and governance will also present significant challenges. In a world where Agents handle sensitive information, and even execute regulated processes (such as in pharmaceuticals or banking), companies must be able to audit and track every task completed by Agents. Long-running Agents may need independent identities to log into various systems, tightly restrict the operations they can perform, and the data they can access. We will need a whole new set of software and platforms to address these issues, much like we have built a security system for human users and applications in the past.
Overall, we are entering a new era of software. In this era, software must be designed from the ground up for massive-scale Agent use. As trillions of Agents work on behalf of humanity, our relationship with software will also be completely reshaped.
You may also like

The large models in the United States are moving towards closure in the name of security

Morning Report | CoinEx becomes a key hub for Iran to evade sanctions, involving over $3.8 billion in funds; Kalshi seeks a new round of financing, with a valuation potentially rising to $40 billion

From the white-haired stock god to the billionaire fund mogul, the smart people shorting Nvidia are all getting rich using the same framework

Why do cryptocurrency projects always like to change their names?

Global Launch: As predictions become the most scarce asset in the AI era, Manadia is defining the next generation of the value internet

Who is footing the bill for the $64 billion accounting frenzy?

I never expected that the first application of AI x Crypto would be in security auditing

What is your view on Binance's competitive advantages?

ETH has entered a non-consensus phase, and the turning point is approaching!

The shift in the cloud of the air: from despising stablecoins a year ago to the high-profile entry of capital today

The survival dilemma of small and medium exchanges behind the withdrawal anomalies exposed by AscendEX

Why Is Bitcoin Falling Below $60K? 5 Key Market Drivers Explained
Bitcoin has dropped sharply amid ETF outflows, Strategy stock weakness, AI stock rallies, and changing Fed expectations. Explore the key forces driving BTC’s latest correction and what traders should watch next.

Bitcoin vs. Gold in 2026: Which Asset Performs Better in Different Markets?

Morning News | The draft amendment to the People's Bank of China Law aims to clarify the legal status of digital renminbi; South Korea will transfer about 40 unregistered virtual asset service providers to law enforcement agencies

The cryptocurrency industry has entered the "Show Me" era: merely relying on vision is no longer enough

Interpreting the Ethereum Foundation's new structure: Reaffirming self-sovereignty amid institutional trends

Former SpaceX engineer reconstructs the financial execution system using first principles






