AI Agents Get Smarter and Safer: Two Years of Progress on WorkBench
AI agents have made huge strides in both performance and safety over the past two years. The best agent now completes nearly 90% of tasks and makes harmful mistakes just 2.5% of the time, down from 26%.
The best AI agent on WorkBench in March 2024, GPT-4, could complete 43% of tasks but took unintended harmful actions, such as emailing the wrong person, 26% of the time. Two years later, in June 2026, the top agent, Claude Opus 4.8, finishes 89% of tasks and only takes unintended harmful actions 2.5% of the time.
This progress means AI agents are becoming more reliable for everyday tasks. For example, they can now handle more complex work like scheduling meetings, drafting documents, or managing projects with fewer errors. The fact that safety and capability improved together is especially good news—it means AI can get better at doing things without becoming more dangerous.
If you're curious about how AI agents work, try out the latest version of Claude Opus 4.8. You can test it by signing up at claude.ai and asking it to help with a task, like drafting an email or organizing your calendar. See how it compares to what you're used to!