Anthropic adds computer use AI tool to Claude

Anthropic on Tuesday introduced a new capability that uses computers the way people do, according to the AI startup.The function -- available now in public beta -- is part of the vendor’s large language model Claude 3.5 Sonnet, an updated version of which Anthropic also released on Tuesday along with another updated model, Claude 3.5 Haiku.Computer enables developers to direct Claude to look at a screen, move a cursor, click buttons and type text.Developers can use the new capability to automate repetitive tasks, build and test software and conduct open-ended functions like research, Anthropic said.To do this, developers will use an API built by Anthropic that can understand and interact with computer interfaces.AI agentsThe tool reflects the shift in the generative AI market to AI agents.The move to AI agents was marked by Salesforce introducing large action models and Agentforce, the CRM giant’s agentic AI platform."Agent-based AI is the next big thing, and everyone is trying to figure out how to do it safely," said Futurum Group analyst Dion Hinchcliffe.Anthropic’s new capability can lead to many agent-based scenarios because it enables generative AI technology to use any application on a computer, Hinchcliffe said.The new capability presents new opportunities for enterprises, he added."Most enterprise activities can potentially be automated using AI with direct computer use," Hinchcliffe continued.According to Anthropic, applications such as mobile work management platform Asana, graphic design platform Canva and food delivery service DoorDash used the computer use tool before it was launched in public beta.Similar to RPA, BPAWhile the new computer use format appears unique, it is similar to many robotic process automation (RPA) and business process automation (BPA) and other task automation systems that the business process management market is familiar with, said Forrester Research analyst Rowan Curran."There are plenty of software companies that offer the exact type of capability that they're launching, the ability to basically have a machine interaction with a human interface, so controlling the pointer, taking on a keyboard, things of that nature," Curran said. "There are plenty of tools that already do this out there in the world."For example, IBM has BPA tools that can do the same tasks. Appian also helps vendors automate business tasks."It's yet to be seen whether there is some specific advantage to how they are operating this capability, and how they might be specifically leveraging it with their family of language models," Curran said.AI computer use may be helpful and important in building larger-scale and complex AI agents across legacy systems where RPA has played a role over the past few decades, he added.Some risksWhile there is significant potential for enterprises looking to use this new tool, it also comes with risks."I don't think it's smart to release yet,” Hinchcliffe said. "I don't think companies are ready to control all the issues ... around security, safety, responsible usage."The capability could make it easy for bad actors to hack systems easily because they direct an AI system to log into a computer system to try to find its vulnerabilities.While some enterprises might move quickly to try the new tool, others will take a careful approach."Smart organizations are going to wait and see if the guardrails are in place," Hinchcliffe said. "There'll be a lot of people who will want to use it right now, and it's such a powerful tool that can cause a lot of problems inadvertently."The computer use function is still imperfect, Anthropic said. The tool currently finds it challenging to scroll, drag and zoom -- tasks that people can do effortlessly.Meanwhile, Claude 3.5 Haiku is strong on coding tasks and can follow instructions better than previous versions, Anthropic said.Claude 3.5 Haiku will be available later this month across Anthropic's first-party API, Amazon Bedrock and Google Cloud's Vertex AI.Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.