Anthropic has unveiled a groundbreaking capability for its Claude large language model: the ability to directly interact with a computer. This new feature, called “computer use,” enables AI agents to move cursors, click buttons, and type text, effectively automating on-screen tasks that previously required human intervention. In an interview with MIT Technology Review, Anthropic’s chief scientist Jared Kaplan highlighted this development as a major step in agentic AI.
The computer-use feature is being tested by companies like DoorDash, Canva, and Asana, signaling its potential to streamline workflows and boost productivity. Kaplan explained that this evolution allows AI to go beyond text interactions, performing real-world tasks such as filling out forms or managing online research. By integrating reasoning and feedback mechanisms, these agents can handle high-stakes decisions while asking users for guidance when needed.
Kaplan also discussed how these agents could transform industries, particularly coding. “Claude will not just autocomplete,” he said, “but debug, run, and even fix code.” Such advances could make AI an invaluable tool for developers and other professionals, fostering collaboration and innovation.
However, Kaplan cautioned about the risks. “Prompt injection attacks, where malicious prompts trick AI into unintended actions, are a significant concern,” he said. Anthropic is focusing on safety as agents become more integrated into workflows, addressing both ethical and technical challenges. Despite the hurdles, Kaplan expressed optimism about the role agents will play in 2025, saying, “AI is accelerating us internally at Anthropic, and the future looks very exciting.”