AI agents have become one of the clearest directions for the next phase of artificial intelligence. The idea is simple in outline but difficult in practice: instead of only answering questions, a system can go away, use tools, and complete tasks with little or no supervision.
That shift is why companies including Google DeepMind, OpenAI, and Anthropic are working to extend large language models beyond chat. Nvidia and Salesforce are also talking about agentic AI as a force that could change how the industry works. Sam Altman wrote that in 2025, the first AI agents may “join the workforce” and materially change company output.
From chatbots to systems that act
An agent, in the broadest sense, is software that does something on a user’s behalf. The more complex the task, the more capable the system has to be. A simple agent might fill in a form, search for a recipe and add ingredients to an online grocery basket, or prepare a short briefing before a meeting.
The important change is that the model is no longer limited to producing text. It has to interpret an environment, choose an action, use a tool, and often check whether the result matches the user’s goal. That makes agentic AI more useful, but it also makes failures more consequential.
Anthropic has already shown one version of this direction through computer use, an extension of Claude that lets the model use a computer in a way that resembles a person: moving a cursor, clicking buttons, and typing. The feature is available to a small group of testers, including third-party developers at DoorDash, Canva, and Asana.
Anthropic says computer use is still cumbersome and error-prone. Even so, it points to the larger ambition: AI systems that can work across ordinary software interfaces rather than only inside a narrow prompt-and-response window.
Better agents need better tool use
Jared Kaplan, Anthropic’s cofounder and chief scientist, frames progress along two axes. One is the complexity of the task an AI system can handle. The other is the range of tools or environments it can operate within.
That distinction matters because intelligence alone does not make a system broadly useful. Kaplan points back to AlphaGo, which showed superhuman performance in board games but remained inside a very restricted environment. Text models, multimodal models, computer use, and perhaps robotics all move AI toward more varied situations.
For agents, that means the next improvements are not only about deeper reasoning. They are also about practical interaction: opening applications, navigating interfaces, recognizing mistakes, and knowing when a task is high-stakes enough to ask the user for feedback.
This is where computer use becomes more than a demo. If Claude can improve at handling different tasks and more complex tasks on a computer, the model becomes closer to a general assistant for digital work. But that usefulness depends on reliability, not just autonomy.
Context will decide whether agents are useful
Kaplan also emphasizes that agents need to understand the user’s situation. A useful Claude would need to learn enough about a person’s role, writing style, organizational needs, and working constraints to act in ways that match expectations.
That could involve searching through documents, Slack, and similar sources to learn what is relevant. In this view, context is not just a productivity feature. It is also part of safety, because a system that understands what the user actually expects is less likely to take the wrong action.
Not every job for an agent requires heavy reasoning. Opening Google Docs, for example, should not require a long chain of thought. Kaplan expects progress in applying reasoning when it is useful and important, while avoiding unnecessary delay when the task is straightforward.
That is a practical point for workplace AI. Users do not only want smarter systems. They want systems that know when to move quickly, when to think more carefully, and when to pause for confirmation.
Coding assistants may become more collaborative
Developers are already a major early audience for these systems. Anthropic released an initial beta of computer use to developers so it could collect feedback while the system was still primitive. As the technology improves, Kaplan expects broader collaboration across different activities.
DoorDash, the Browser Company, and Canva are experimenting with browser interactions and designing them with help from AI. That shows how agents may fit into workflows where software has to be operated, tested, and adjusted rather than merely described.
Coding assistants are another area where Kaplan expects further gains. He notes strong interest in using Claude 3.5 for coding, with systems moving beyond autocomplete. The more advanced pattern is an assistant that can understand what is wrong with code, debug it, run it, observe what happens, and then make a fix.
That is a meaningful change in the role of AI in software work. The assistant is not only suggesting the next line. It is participating in a loop of diagnosis, execution, and correction.
Safety becomes harder as agents gain access
The same qualities that make AI agents useful also raise safety concerns. A model that can browse, click, type, and act across software has a larger surface for mistakes and misuse than a model that only replies in text.
Kaplan identifies prompt injection as one of the central problems for wider agent use. The source describes prompt injection as an attack where a malicious prompt reaches a large language model in an unexpected or unintended way, including through websites that models may visit.
That risk is especially important for computer use. If the feature were deployed at large scale, harmful websites could try to persuade Claude to do something it should not do. Anthropic says it is working actively on this issue.
Kaplan also connects agent safety to more advanced models more generally. Anthropic has a robust scaling policy, and he says that as AI systems become sufficiently capable, the company needs to prevent misuse, including cases where systems could help terrorists.
The central tension is clear. AI agents may become more embedded in work, especially in coding and other digital tasks. But the closer they get to real action, the more important it becomes that they understand context, detect danger, ask for help when needed, and stay within the user’s intent.