A major reshuffle in the software industry: When the target users are no longer humans but trillions of intelligent agents

God Translation Bureau is a compilation team under 36Kr, focusing on fields such as technology, business, workplace, and life, with a particular emphasis on introducing new technologies, new ideas, and new trends from abroad.

Editor's note: Is software no longer developed for humans? When the number of agents reaches a thousand times that of humans, APIs and pay - per - use models will become the only survival rules for developers in this era. This article is a compilation.

In the past few months, there have been significant changes in the field of agents. At the end of last year, we entered a stage where coding agents can handle tasks with longer operation cycles and no longer require excessive human intervention throughout the development process.

These agents are no longer just chatbots with basic tools. Instead, they usually have independent sandbox computing environments, can write and run code to solve any problems they encounter, directly interact with APIs and command - line interfaces (CLI), and have their own file systems and long - term memories. The popularization of this set of core primitives and best practices in agent frameworks, along with the remarkable progress of models in tool invocation and software development, allow us to glimpse the future of agents that can handle any task.

Although this architecture was initially defined by coding agents such as Claude Code, Devin, Codex, Factory, Cursor, or Replit, we have recently crossed the gap and entered various fields of personal experience and knowledge work. This is thanks to the emergence of agents such as Claude Cowork, Perplexity Computer, Manus, and OpenClaw. In particular, OpenClaw has pushed the future even further by implementing an agent that runs 24/7 in a persistent environment.

With the rapid improvement of their capabilities, agents will penetrate into almost all work fields. They will be deployed to review every contract, handle the vast majority of front - line customer support, audit the finances of every company, sort out every medical research for drug development, generate almost all code, create most sales and consulting presentations, and represent consumers in online transactions. In short, agents will participate in almost every economically valuable task in society.

This is not just about performing our existing tasks. We will use agents to complete work far beyond what we have done before - running simulations that were previously unaffordable and prototyping multiple solutions for each creative idea. Due to the low startup costs and ease of shutdown, we will launch more projects and conduct a comprehensive review of every piece of data instead of just sampling.

Overall, we can foresee that each employee in an organization will have multiple agents working for them. It is not difficult to imagine that the number of agents in a company will be 100 times or even 1000 times the number of employees. As trillions of agents work and collaborate, agents will become the main users of all future software.

Given that most software is designed for humans, this means we will witness a major transformation in the future form of software. So, what's next?

Build What Agents Want

Paul Graham once summarized how to develop software in the most concise way: Build what people want.

This advice has contributed to some of the greatest software success stories in the 21st century and has driven a movement to build tools that are easy to use, easy to adopt, can solve clear problems in plain language, and have clear and transparent pricing.

Now, the direction forward is to develop software that agents want to use. Although the largest user group of agents currently consists of developers or highly technical users who usually have their own tool preferences, in a world where agents perform various tasks for knowledge workers, these preferences will gradually fade. Unless a company already has a standard, agents will dominate the adoption of tools in specific workflows.

This may involve the tools they register for, the code they write, the libraries they use, the skills they leverage, and so on. Platforms that are more easily adopted by agents and can perfectly solve the problems of agents (and users) will progress much faster than those that cannot. Agents won't attend your webinars or look at your ads; they will simply choose the tools that are most suitable for the task, and you definitely want that tool to be yours.

The most important implication of this advice is that everything you build must be "API - first". If you don't provide an API for a feature, it might as well not exist. If it can't be exposed through a command - line interface (CLI) or a model context protocol (MCP) server, you're at a disadvantage. If your API is confusing or provides conflicting execution paths for agents, you're reducing your value to agents. When building a file system for agents, we've been sorting out all aspects of the API to identify where things might break down in the agent world and achieve a level of usability that was previously only considered in user experience (UX) design.

Just as designing software for users means putting yourself in their shoes, thinking about what agents will encounter is the same. For example, Jared Friedman of Y Combinator once reminded everyone: "Even the best developer tools mostly still don't support account registration via API. This is a major oversight in the era of Claude Code, because it means Claude can't register on its own. Connecting all account management functions to the API should be a basic requirement now." If agents can't easily register and start using your service, you're basically out of the game in the eyes of agents.

In a world where agents become the largest users of future software, business models will also face major adjustments. In some cases, agents triggered by user seats may still be suitable for seat - based business models, but there are many agent use cases that cannot be linked to existing users, or because their workloads are completely different. For example, an agent can complete the work equivalent of several hours of human work in software with just a few sentences or lines of text and only present the final result to the end - user.

This will ultimately mean the evolution of some software business models, because any tool that wants to survive in the agent era needs to have some form of pay - per - use or scale - based business model built into its system, and may even need to support agents being able to pay for these services on their own.

The Next - Generation Infrastructure and Tools Will Be Built for Agents

It was a good idea to give computers to humans. But it's an even better idea to give computers to computers so that they can create the same output on computers as we do in our work.

—— Aravind Srinivas, Perplexity

As agents have their own computers, can write and execute their own code, call common skills for repetitive operations, and access external tools and services, this will create opportunities for a whole set of new technologies dedicated to agents. Imagine what users do on their computers, and agents will need a similar set of capabilities designed specifically for them.

Some of these core services will naturally come from existing players, as agents are using existing data, or the collaboration or connection between existing human users and agent users in the system has value. Similarly, new categories will also emerge because these problem spaces are so different from the previous needs or capabilities of human users that it makes sense to design services from scratch.

For example, it's obvious that agents need their own infrastructure to run, and the scale will be unprecedented. The next hyperscale cloud service provider (or an existing one) will be built on the idea that future server farms will serve our agents rather than our applications. E2B, Daytona, Modal, and Cloudflare are all working in this direction, and these sandbox environments will challenge any computing scale we've seen before.

Agents also need to access the core files of enterprises and manage their own memories and long - term work data, which is what we're committed to building. Similarly, large - scale enterprise systems also need to be "API - first" so that agents can handle key services and data in the organization, such as human resources information systems (HRIS), customer relationship management (CRM), workflows, data lakes, and other major systems. Products that can provide the most seamless tools for agents to operate this data anywhere will be in the best position to win these future workloads.

Agents may also need identity identifiers and the ability to communicate with other parties. For example, Agentmail is providing mailboxes for agents so that they can have their own persistent emails. Companies like Parallel and Exa are rebuilding web search to adapt to a world where agents are the largest users of web information scraping. Many types of agents will need to manage their expenditure budgets through wallets like Stripe or Coinbase, and we may finally see the real - world application of micropayments, through which agents can access paid tools and information.

Security, compliance, and governance will be the main issues faced by these agents. In a world where agents access and process sensitive information in workflows or perform regulated workflows (such as in the pharmaceutical or banking industries), companies need to govern and retain all the work done by these agents. Long - running agents may need their own identities to authenticate into various services and have strict control over the types of operations they can perform and the data they can access in the enterprise. We need new software and platforms to address these challenges, just as we've built for humans and applications over the years.

Overall, it's obvious that we're entering a new era of software, and we need to design and build tools specifically for the large - scale use of agents. In a world where trillions of agents perform work, this will open up a whole new way of software collaboration.

Translator: boxi.