2025 is being the year in which artificial intelligence has stopped being a tool that waits for orders to become an autonomous agent capable of acting on its own. If 2023 was the year of chatbots (ChatGPT, Gemini, Claude) and 2024 the year of multimodal models that saw and heard, 2025 is undoubtedly the year of the “AI Agents”: assistants that not only answer questions, but also browse the internet, fill out forms, write code, operate programs and execute complex tasks without constant supervision.What is an autonomous AI agent?Unlike a traditional chatbot that responds to a prompt and stops, an autonomous agent receives a broad goal (“organize a trip to London for three people” or “analyze these 50 PDFs and generate a summary”) and breaks it down into steps, executes each one, evaluates results and corrects course if something fails. It is like having an indefatigable digital intern.The protagonists of 2025
- Anthropic Claude with Computer Use: Claude can see the computer screen, move the mouse, click buttons and type text in any application. It does not need an API or special integration: it interacts with software just like a human. The 2025 update drastically improved its accuracy when clicking on the correct elements.
- OpenAI Operator: OpenAI’s assistant that browses the web autonomously. It can fill out complex forms, make online purchases, book restaurants and manage administrative tasks. Launched early in the year, it has been improving its ability to handle CAPTCHAs and two-factor authentication.
- Google Project Mariner: Google’s bet on a Gemini-based agent that operates inside Chrome. It can research topics, extract data from multiple web pages and fill in spreadsheets automatically.
- DeepSeek Agent: the Chinese model DeepSeek V4 (released in July 2025) includes native agent capabilities, allowing deep reasoning + execution of actions in a single flow, competing directly with the Western giants.
Practical applicationsAutonomous agents are revolutionizing entire sectors:
- Customer support: agents that manage complete incidents (opening a ticket, investigating, responding and closing) without human intervention.
- Software development: tools such as Claude Code or Cursor Agent write code, run tests, make commits and deploy to production.
- Market research: agents that browse dozens of sources, extract data and generate executive reports.
- Process automation: from invoice management to inventory updates.
Lights and shadowsNot everything is advantages. The ability of agents to act on the internet poses important security risks: what happens if an agent makes a mistake and deletes important files? who is responsible if an AI makes an unauthorized purchase? European regulators are already studying how to fit these agents into the framework of the EU AI Act. In addition, the computational consumption of agents is much higher than that of traditional chatbots, which makes their massive use more expensive.The immediate futureExperts predict that by the end of 2025 we will have AI agents specialized by industry: legal agents that manage documentation, medical agents that prepare clinical reports, financial agents that monitor markets. Interoperability between agents from different providers (so that an Anthropic agent can call a Google agent) will be the next great technical challenge. What is clear is that the era of “AI that only chats” has ended. Welcome to the era of AI that does things.
