This chart comes from Anthropic’s new research paper, “Measuring AI agent autonomy in practice”, published this week. The team analysed millions of human-agent interactions across both Claude Code and Anthropic’s public API, studying individual tool calls as the building blocks of agent behaviour. The result? Software engineering accounts for 49.7% of all agentic activity. The next biggest category, back-office automation, manages just 9.1%. Everything else is in the low single digits. Customer service, a domain that has been talked about as a natural fit for AI agents for years, sits at just 2.2%, behind data analysis, cybersecurity, and even academic research.
The research also revealed that the longest autonomous Claude Code sessions nearly doubled in three months, from under 25 minutes to over 45 minutes. Experienced users auto-approve more actions but also interrupt more often, shifting from step-by-step approval to active monitoring. And Claude itself pauses to ask for clarification more than twice as often as humans interrupt it on complex tasks.