They discovered that while most of the usage is for one-shot code snippets, more users are letting Claude Code work autonomously, up to 45 minutes at a time after three months.
More people are auto-approving code, but he more experienced users tend to interrupt the workflow more often.
Asks for clarification
Additionally, Claude Code pauses for clarification more than it gets interrupted — at twice the rate. That means users are more likely to see a clarification prompt from Claude than they need to actually intervene.
Anthropic also found most interactions with Claude Code to be low risk. 50% of the agent usage is for software engineering tasks, which can be from full programming to fixing bugs, but with little risk to lives and limbs.
They expect more risk from other usage patterns, like when agent use moves into the realms of financial transactions and medical information, which currently account for a usage of 4% and 1% respectively.
Read the full study at Anthropic, and see the summary thread.