Original Title: Building for trillions of agents
Original Author: Aaron Levie, Box
Translation: Peggy, BlockBeats
Editor's Note: As the capabilities of large models continue to break through, AI Agents are gradually evolving from "dialogue tools" to digital labor forces capable of executing tasks independently. From writing code and handling contracts, to auditing finances and analyzing research data, Agents are beginning to enter almost every aspect of knowledge work.
When the number of Agents operating within a company far exceeds the number of employees, the primary users of the software may shift from "humans" to "machines." Under this trend, software design, infrastructure, and even business models are changing. This article discusses how the software forms and infrastructure of the Agent era will evolve, using "building software for Agents" as the thread.
Note: The author of this article, Aaron Levie, is a co-founder and CEO of the enterprise cloud storage company Box, and a long-time technology industry opinion leader focusing on AI and enterprise software trends.
The following is the original text:
In the past few months, an important change has been taking place in the field of Agents. Around the end of last year, we began to enter a stage where programming Agents are capable of independently completing longer-duration tasks without the need for frequent human guidance throughout the development process.
These Agents are no longer just chatbots equipped with simple tools. Nowadays, they often have independent sandbox computing environments, can autonomously write and run code in response to encountered issues, can directly call APIs and CLIs, interact with various systems, and possess their own file systems and long-term memory capabilities, among other features. These foundational abilities, combined with the gradual maturity of best practices surrounding the Agent operating framework (agentic harness), and significant advancements in models for tool invocation and software development, lead us to see a possibility: Agents will be able to handle almost any task assigned to them.
Initially, this architecture was primarily driven by a group of programming Agents such as Claude Code, Devin, Codex, Factory, Cursor, Replit, among others. But recently, this model has crossed early tech circles and begun to enter wider personal experiences and knowledge work domains, like Claude Cowork, Perplexity Computer, Manus, and of course, OpenClaw. The latter has pushed this direction further—operating continuously in a persistent environment for 24 hours.
As capabilities rapidly improve, Agents will be introduced into almost all fields of work. They will be used to review every contract, handle a large volume of frontline customer support issues, audit corporate finances, sift through vast amounts of medical research to drive drug discovery, generate most software code, create sales and consulting presentations, and even perform transactions on the internet on behalf of consumers. Overall, they will participate in nearly all economically valuable work in society.
Moreover, this does not merely mean completing tasks we are already doing today. Agents will enable us to do more, such as run complex simulations that have previously been too expensive, quickly generate multiple prototype solutions for every idea due to drastically lower project initiation costs, making stopping projects easier; we will push forward more projects simultaneously; and we can analyze nearly all data rather than just relying on sampling.
Looking at these trends collectively, it is foreseeable that in future organizations, nearly every employee will have multiple Agents working for them. It is not hard to imagine a company having a number of Agents that is 100 times or even 1000 times the number of employees. When trillions of Agents operate simultaneously, they will become the primary users of future software.
However, most software was originally designed for humans. This means that software forms are likely to undergo a significant change. So what will happen next?
Creating Software That Agents Are Willing to Use
Paul Graham once summarized the principle of software entrepreneurship in an extremely simple way: make something people want.
This idea has spawned some of the most successful software companies of the 21st century and driven a new product methodology—tools should be simple to use, easy to get started with, solve clear problems, avoid obscure terminology, and have transparent pricing.
Now, this phrase may need to be rewritten as: create software that Agents are willing to use.
Currently, the people who use Agents the most are often developers or technically proficient users who have their own preferences for tools. However, as Agents begin to handle various tasks for knowledge workers, this human preference will gradually weaken. Unless a company has already established unified tools internally, in many workflows, the real decision-maker will be the Agent.
This means that they will determine which tools to use, what code to write, which libraries to call, and which skills to apply. Platforms that are easier for Agents to access and that better solve problems will gain advantages more quickly than other products. Agents will not attend your online launch events, nor will they see your ads; they will only choose the tools that are most effective for completing tasks, and you will certainly hope that those tools are your products.
The greatest insight from this advice is that everything must be API-centered (API-first).
If a feature does not have an API, it is almost equivalent to not existing.
If a feature cannot be called through CLI or MCP server, you are already one step behind.
If API design is complex and chaotic, with path conflicts that make it difficult for Agents to use, you are basically actively giving up the opportunity to become an Agent tool.
At Box, we are committed to building a file system for Agents, so we are examining every detail of our API, considering where problems may arise in an Agent environment. This level of detail has often only been observed in user experience (UX) design in the past.
Just as designing software for users requires thinking from the user’s perspective, designing software for Agents requires the same mindset. For example, Jared Friedman from Y Combinator reminds developers: "Even the best developer tools still require most to register accounts through APIs. This is a huge issue in the Claude Code era because it means Claude cannot register an account by itself. Now, putting all account management features into the API should become a basic requirement."
If Agents cannot easily register and start using your service, then in their eyes, you are almost equivalent to non-existent.
Business Models Will Change Accordingly
When Agents become the primary users of software, business models will also undergo changes.
In some cases, a model triggered by user seats (seat-based) for running Agents is still applicable. But there will also be many Agents that are no longer tied to specific users, or their workload will far exceed traditional software usage patterns. For example, a user may simply input a few sentences, and an Agent could complete hours of human work within the software, only presenting the final result to the user.
Therefore, the business models of some software products will evolve. Any tool hoping to survive in the "Agent era" will need to introduce some sort of usage-based or computation-based billing model, and even support Agents to complete payments independently.
Next-Generation Infrastructure for Agents
Aravind Srinivas, the founder of Perplexity, once said: "It is a good idea to hand computers over to humans, but handing computers over to computers to complete work for us is a better idea."
As Agents possess their own computing environments, can write and execute code, call skills to complete repetitive tasks, and connect with various external tools and services, a whole new technical system will emerge as well. Just as humans need on computers, Agents also need a similar but specifically designed infrastructure.
Some 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, numerous new product categories will emerge, as the problems Agents face are entirely different from those faced by humans, often making it more rational to design new services from scratch.
For example, Agents clearly need their own infrastructure, and the scale may be unprecedented. Next-generation ultra-scalable cloud platforms (or upgraded versions of existing giants) could likely be built around the concept that future data centers will no longer run our applications, but run our Agents. Companies like E2B, Daytona, Modal, Cloudflare are already moving in this direction, and the scale of these sandbox computing environments may reach previously unseen levels.
Agents will also need access to core corporate documents and manage their own data and memories to support long-running tasks. Similarly, enterprise systems will need to transition to API-first, so that Agents can access critical data and services from HR systems, CRMs, workflow systems, data lakes, etc. Platforms that enable Agents to seamlessly operate this data anytime and anywhere will be most likely to bear future workloads.
Agents may also need their own identity systems and be able to communicate with each other. For example, Agent mail is providing Agents with email addresses to have a persistent email address. Meanwhile, companies like Exa and Parallel are also reconstructing search engines to adapt to a world where "Agents are the primary search users." Many Agents will also need to manage their own budgets, for example, using wallets provided by Stripe or Coinbase to complete payments, which may even drive the realization of micropayments, allowing Agents to access paid tools and data resources.
Of course, security, compliance, and governance will also become significant challenges. In a world where an Agent handles sensitive information or even executes regulated processes (like pharmaceuticals or banking), companies must be able to audit and log every task completed by the Agent. Long-running Agents may need independent identities to log into various systems, and strictly limit their operations and the data they can access. We will need an entirely new set of software and platforms to address these issues, just as we built security systems 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 outset for large-scale use by Agents. As trillions of Agents work for humans, our relationship with software will also be fundamentally reshaped.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
