Measuring AI Agent Autonomy in Practice
Anthropic analyzes millions of Claude Code and API interactions to see how much autonomy users give agents and where risks show up.
Intelligence analysis by GPT-5.4 Mini
The post studies real-world agent use across Claude Code and Anthropic's public API. It finds that users hand over more work as they gain experience, Claude Code runs longer without intervention, and most agent actions are still low-risk and reversible.
Anthropic watched how people use its AI helper. It found that, with practice, people let the helper work more on its own, like giving a bike rider more freedom after they learn. Even then, the helper sometimes stops to ask a question when the path looks tricky.
Analysis
Method
Anthropic defines an agent as an AI system with tools that let it take actions, such as running code, calling external APIs, or messaging other agents. To study how agents behave in practice, the company combines two sources: its public API, which gives broad visibility across many customers at the level of individual tool calls, and Claude Code, which lets Anthropic trace full sessions and measure autonomy over time.
Findings
One main result is that Claude Code is staying active without human intervention for longer stretches. In the longest-running sessions, the time before Claude stops has nearly doubled in three months, from under 25 minutes to more than 45 minutes. Anthropic says that increase is smooth across model releases, which suggests the change is not just a side effect of newer models getting better. The company reads this as evidence that existing models may already support more autonomy than people usually let them use.
Experience also changes behavior. New users often keep reviewing every action, while experienced users more often switch to full auto-approve and intervene only when needed. The post says roughly 20% of sessions use full auto-approve among new users, rising to over 40% as users become more familiar with the tool.
A second finding is that the agent itself sometimes applies the brakes. On complex tasks, Claude Code pauses to ask for clarification more often than humans interrupt it. Anthropic treats those agent-initiated stops as a meaningful layer of oversight, not just a nuisance.
On the public API, most actions are described as low-risk and reversible. Software engineering makes up nearly half of agentic activity, but the company also sees emerging use in healthcare, finance, and cybersecurity. The article closes by arguing that wider deployment will require better post-deployment monitoring and new human-AI interaction patterns that manage autonomy and risk together.
Key points
- Claude Code is running for longer without human interruption, with the longest sessions nearly doubling over three months.
- Experienced users are more likely to use full auto-approve and less likely to review every action.
- Claude Code often stops to ask for clarification on hard tasks before humans interrupt it.
- Most public API activity is still low-risk and reversible, but use is starting to appear in higher-stakes domains.
- Anthropic says better post-deployment monitoring and new oversight patterns will be needed as agent autonomy grows.
If the trend continues, people may be able to hand off longer, more useful work while still getting help when the AI is unsure. The article also suggests that many current actions are low-risk and reversible, which could make broader use safer as monitoring improves.
As users become more comfortable, they may check less often, which makes mistakes harder to catch early. The post also notes growing use in healthcare, finance, and cybersecurity, where even small errors can have bigger consequences.
