Think AI agents are working autonomously for hours on end? Anthropic has real data to show otherwise — most tasks are done in 45 seconds.
But another figure is more noteworthy: the longest sessions are growing twice as fast.
Research Methodology
Anthropic analyzed millions of real interactions on Claude Code and its API to understand three things: how much autonomy users grant agents, in which scenarios agents are deployed, and what potential risks exist.
This is the first time the industry has used large-scale real-world usage data to measure AI agent autonomy, rather than relying on benchmarks or sandbox experiments.
Key Data Points
Median session duration: 45 seconds
The vast majority of agent tasks are short and quick. These are not multi-hour complex projects, but specific, small tasks.
However: Between October 2025 and January 2026, the 99.9th percentile session duration in Claude Code grew from under 25 minutes to over 45 minutes. In other words, the longest 0.1% of tasks doubled in three months.
This trend indicates that the heaviest users at the top end are pushing AI agents into increasingly longer and more complex workflows.
Where Agents Are Used
Software engineering accounts for 49.7% of all tool calls. This proportion is intuitive — programmers were the earliest to adopt agents at scale, and Claude Code was designed for this scenario.
But the figures below are more revealing:
| Scenario | Share |
|---|---|
| Software engineering | 49.7% |
| Backend automation | 9.1% |
| Marketing/copywriting | 4.4% |
| Sales/CRM | 4.3% |
| Finance/accounting | 4.0% |
| Data analysis | 3.5% |
Backend automation, finance, and sales are quietly penetrating. These are not lab demos — real enterprises are using AI agents to handle actual business processes.
Is Human Oversight Still Present?
80% of tool calls have at least one safeguard. 73% involve a human-in-the-loop step. Only 0.8% of operations are irreversible.
These numbers are generally optimistic — they show that most enterprises still retain safety valves when deploying agents.
But user behavior shows an interesting pattern of change:
- New users: about 20% of tasks choose fully automatic approval
- Experienced users (750+ sessions): fully automatic approval exceeds 40%
This looks like humans are ceding control as trust increases. However, the interruption rate is also rising — from 5% to 9% over the same period.
This suggests experienced users are not blindly letting go, but are shifting their oversight style: from approving each action step-by-step to overall monitoring and intervening only when problems arise.
AI Agents Also Pause Themselves
An unexpected finding: Claude Code requests clarification on complex tasks more than twice as often as on simple tasks, and this self-initiated pause frequency is higher than the rate of human manual interruptions.
In other words, the model itself can identify moments when it is uncertain and needs to ask. This is more granular than relying solely on human oversight.
Deployment Overhang
The study introduces a concept: deployment overhang.
This means that the current level of autonomy granted to models is actually below their capability ceiling. Over time, users will gradually push tasks to the boundary — longer sessions, less intervention, more complex task chains.
The implication for safety is that pre-release testing alone is insufficient. There is a need to establish continuous monitoring infrastructure after deployment. Just as driving relies not only on factory inspections but also on dashcams and traffic enforcement.
Sources: CocoLoop, Measuring AI agent autonomy in practice (Anthropic Research); The Autonomy Gap: What Anthropic Learned Watching Millions of AI Agent Interactions (Shashi.co)