AI agents are becoming more common in practice, but their use is still heavily focused on programming, according to new analysis by Anthropic of millions of interactions between humans and AI agents, where researchers studied how tools like Claude Code and public API solutions are used in real-world environments.
According to the report, software development accounts for nearly half of all agent-based activity. AI agents are primarily used to write, run, and test code, and users are increasingly letting them work autonomously for longer periods of time. Among the longest sessions, the amount of time the agent worked autonomously almost doubled in three months, from under 25 minutes to over 45 minutes.
At the same time, its use outside of programming is still limited. While researchers see early examples in areas such as healthcare, finance, and cybersecurity, these are still small-scale and mostly involve low-risk tasks that can be undone or corrected.
The researchers believe that developments point towards a future where AI agents become increasingly autonomous, but that broader use outside of programming likely requires new ways of monitoring and controlling the interaction between humans and AI.