Anthropic announced two new AI agents that it claims are a major step towards making AI agents useful.
The company claims that AI agents trained on Claude Opus 4 – the company’s most powerful AI model – raise the bar in terms of what these systems are capable of. They can tackle difficult tasks for extended periods of time, and respond more effectively to user instructions.
Claude Opus 4 is designed to perform complex tasks that require thousands of steps completed over several hours. It created a guide to the video game Pokemon Red by playing it for over 24 hours straight. Dianne Penn, Anthropic’s product lead for research, says that the company’s previous most powerful model, Claude 3.7 Sonnet was only capable of playing for 45 minutes.
Similarly, the company claims that one of its clients, the Japanese technology firm Rakuten, deployed Claude Opus 4 recently to code autonomously on a complex open-source project for close to seven hour.
Anthropic made these improvements by improving the model’s ability to create “memory” files to store important information. This improved ability to “remember”makes the model more efficient at completing long tasks.
Penn says that this generational leap is a shift from an assistant to a real agent. While you still need to provide a lot of feedback in real-time and make all the key decisions for AI Assistants, an agent could make these key decisions themselves. It allows humans to be more like a delegator, or judge, instead of having to hold the systems’ hands at every step.
Claude Opus 4 is only available to Anthropic’s paying customers. Claude Sonnet 4 will be available both for paid and free users. Opus 4 is marketed as an impressive, large model that can handle complex challenges. Sonnet 4 is marketed as an efficient, smart model for everyday use.
Both models are hybrids, which means they can respond quickly or more thoroughly, depending on the nature and content of a query. Both models can improve their output by searching the web or using other tools while they calculate a reply.
AI companies are in a race right now to create useful AI agents capable of planning, reasoning, and executing complex tasks reliably, without human supervision. Stefano Albrecht is the director of AI for DeepFlow, a startup, and coauthor of Multi-Agent Reinforcement Learning: Modern Approaches and Foundations. This often involves using the internet or another tool autonomously. There are still many safety and security issues to be overcome. AI agents powered by large-scale language models may act erratically or perform unintended acts. This becomes a problem if they are trusted to act without supervision.
“The more agents that are able to do something over a long period of time, the better they will be if I don’t have to intervene as much,” he says. “The ability of the new models to use tools simultaneously is interesting. That could save time along with the way. So that’s going be useful.”
Agents can take unexpected shortcuts or exploit loopholes in order to achieve the goals they have been given. They might book all seats on a flight to make sure their user gets a place, or resort creatively to cheating in order to win a game of chess. Anthropic claims it has reduced this behavior (known as reward hacking) in both models by 65% compared to Claude Sonnet 3.7. This was achieved by closely monitoring problematic behavior during training and improving both the AI’s training environment and its evaluation methods.
