Anthropic published this study on 18 February 2026 as a first empirical look at how people actually deploy and use AI agents. It analyzed millions of human-agent interactions using the company’s privacy-preserving analysis tooling, asking how much autonomy people grant agents, how that changes as they gain experience, which domains agents operate in, and whether the actions agents take are risky.
For the study, an agent is defined as an AI system equipped with tools that allow it to take actions, such as running code, calling external APIs or sending messages to other agents. The study draws on two data sources with opposite trade-offs. On the public API it analyzes individual tool calls across thousands of customers, which gives breadth but cannot reconstruct how actions compose into longer sessions. In Claude Code, Anthropic’s own coding agent, it can link requests into whole sessions, which makes it the main source for measuring autonomy, but it covers only one product. Much of the classification — of risk, autonomy, task complexity, human involvement and the reasons for stopping — was itself done by Claude, validated against internal data where possible.
Findings
- Claude Code works autonomously for longer. Most turns are short: the median lasts around 45 seconds and has stayed between 40 and 55 seconds. At the tail, the 99.9th-percentile turn duration nearly doubled between October 2025 and January 2026, from under 25 minutes to over 45 minutes. The increase was smooth across model releases, which the authors read as a sign that it is not purely a result of increased capability. Among Anthropic’s internal users, the success rate on the most challenging tasks doubled from August to December while average human interventions per session fell from 5.4 to 3.3. The authors take both measurements as pointing to a Deployment overhang: the autonomy models can handle exceeds what they exercise in practice.
- Experienced users auto-approve more, but also interrupt more. Newer users (fewer than 50 sessions) use full auto-approve in roughly 20% of sessions, rising to over 40% by 750 sessions. Interrupt rates rise too, from about 5% of turns for users with around 10 sessions to around 9% for more experienced users. The study interprets this as a shift in oversight strategy from approving each action to letting the agent run and stepping in when something goes wrong. On the public API, 87% of tool calls on minimal-complexity tasks involve some human, against 67% on high-complexity tasks.
- Claude Code pauses for clarification more often than humans interrupt it. On the most complex tasks it asks for clarification more than twice as often as on minimal-complexity tasks, and more often than humans interrupt it. The most common reasons it stops itself are to present a choice between approaches (35%) and to gather diagnostic information or test results (21%); the most common reason humans interrupt is to provide missing technical context or corrections (32%). The authors caution that Claude may not be stopping at the right moments and that product features such as plan mode may affect the behaviour.
- Agents are used in risky domains, but not yet at scale. In a sample of 998,481 public-API tool calls, 80% came from agents that appear to have at least one safeguard, 73% appear to have a human in the loop in some way, and only 0.8% of actions appear irreversible. Software engineering accounts for nearly 50% of tool calls, with smaller use in areas such as business intelligence, customer service, sales, finance and e-commerce. The riskiest clusters involved security-related actions, financial transactions and medical information, and the authors expect many of them to be evaluations rather than production use.
The study lists its own limitations: it covers a single model provider; API traffic is analyzed as isolated tool calls; the classifications are generated by Claude; it reflects late 2025 to early 2026 only; and tool-call sampling over-represents deployments that make many sequential calls.
Its recommendations are that model and product developers invest in post-deployment monitoring, that model developers train models to recognize their own uncertainty, and that product developers design for user oversight. It argues that it is too early to mandate specific interaction patterns such as requiring humans to approve every action, and concludes that the autonomy agents exercise in practice is co-constructed by the model, the user and the product. See also Human-in-the-loop.