The Rise of Autonomous AI Agents: How Humans Are Adapting to a New Era of Collaboration
Artificial intelligence is no longer simply responding to direct requests. AI agents are now independently undertaking tasks, making decisions within ongoing processes, and executing actions in the real world – from automated software development and customer query triage to financial transactions and medical information services. Recent research from Anthropic sheds light on how people are actually using these systems, the level of freedom they grant them, how behavior evolves with experience, and the resulting risk profiles.
Understanding Agent Autonomy: A Two-Pronged Approach
Anthropic’s analysis, published on February 20, 2026, is based on two key data sources. The first is the general API traffic from Anthropic, providing a broad cross-section of thousands of customer projects. Analysis at this level focuses on individual tool calls, allowing for consistent comparisons across diverse scenarios. The second source is Claude Code, Anthropic’s own coding agent. This allows for a session-by-session linking of requests, reconstructing complete workflows from start to finish. This depth is particularly valuable for studying autonomy – how long an agent works without human intervention, what triggers interruptions, and how user monitoring behavior changes with experience.
For the purposes of this study, an agent is defined as a KI system equipped with tools enabling independent action – including code execution, external interface calls, or communication with other agents. The types of tools used reveal what agents are actually doing.
Longer Autonomy Phases: A Trend Across Models
A central finding of the research concerns the duration of time Claude Code operates without human input. The median “turn duration” – the time between the start and end of a work step – remains stable at around 45 seconds. However, the most revealing signal lies at the upper end of the distribution. The longest working phases, representing the most ambitious applications, have nearly doubled from under 25 minutes (October 2025 – January 2026) to over 45 minutes.
Remarkably, this increase is consistent across all model versions. If greater autonomy were solely a function of improved model capabilities, each new version would show a significant jump. The continuity of the trend suggests multiple factors are at play: experienced users build trust over time, assigning increasingly challenging tasks, and the product itself is continuously evolving. This also suggests existing models are capable of more autonomy than they are currently granted – a potential for increased efficiency.
Internal usage data at Anthropic corroborates this finding. From August to December 2025, the success rate of Claude Code on particularly challenging internal tasks doubled, although the average number of human interventions per session decreased from 5.4 to 3.3.
Experienced Users Monitor Differently – Not Less
As user experience grows, the way people control agents fundamentally changes. Among new users (less than 50 sessions), around 20% of all actions are automatically approved. For users with over 750 sessions, this figure rises to over 40%. This shift reflects a growing level of trust.
Counterintuitively, the rate of manual interruptions also increases with experience. New users intervene in approximately 5% of cases, while experienced users do so in around 9%. This is because beginners approve each step before execution, rarely needing to stop the agent mid-process. Experienced users allow the agent to work independently, intervening only when correction or redirection is necessary. Control, means informed observation with the readiness and ability to intervene.
Similar patterns are observed in API traffic. For less complex tasks, human involvement is evident in 87% of tool calls. For highly complex tasks, this drops to 67%. Complex tasks are disproportionately undertaken by experienced users who grant the agent greater autonomy.
Agents Inquire for Clarification More Often Than Humans Stop Them
Beyond human-initiated interruptions, agent-initiated pauses play a crucial, often overlooked role. Claude Code requests clarification more than twice as often for complex tasks compared to simple ones – and more frequently than users interrupt the agent.
Typical triggers for self-initiated pauses include missing information, unclear requirements, or situations where the agent anticipates consequences beyond expectations. Anthropic views this ability to recognize uncertainty and proactively seek clarification as a vital safety feature, complementing external safeguards like permission systems.
Risk Distribution: Low Averages, Growing Range
To assess risk and autonomy, a KI-based rating system on a scale of 1 to 10 was used. Low risk scores indicate actions with minimal consequences from errors, while high scores represent actions with significant potential impact. The overall picture is reassuring: the vast majority of agent actions on the public API are low-risk. Only 0.8% of actions are considered irreversible, such as sending an email to a customer. Around 80% of tool calls occur within environments with at least one security measure, like restricted access or approval requirements. 73% of cases involve some form of human involvement.
However, clusters with high risk and autonomy values exist at the edges of the distribution, including security-related actions, financial transactions, and medical information processing.
Software Development Dominates, Other Industries Catching Up
Currently, nearly 50% of all tool calls via the public API are related to software development. Other sectors – Business Intelligence, Customer Service, Sales, Finance, and E-commerce – each account for only a few percentage points of total traffic. This reflects the early adoption of agents by software developers. However, data indicates that other industries are beginning to experiment with agent-based solutions. As these areas – particularly healthcare, finance, and IT security – involve higher risk potentials, the boundaries of risk and autonomy are expected to expand.
Recommendations for a Future with Autonomous Agents
Anthropic recommends investing in post-deployment monitoring infrastructure for model and product developers. Developing methods to reliably link independent API requests into coherent agent sessions, while preserving privacy, is a key research area.
Training models to recognize and proactively disclose their own uncertainties is also encouraged. Regulatory requirements should focus on ensuring humans can effectively observe and intervene, rather than imposing rigid, step-by-step approval processes.
autonomy arises from the interplay of model, user, and product. Understanding how agents behave requires observing them in the real world – and the infrastructure for doing so is still in its early stages of development.
FAQ
Q: What is an AI agent?
A: An AI agent is a KI system equipped with tools that allow it to act independently, such as executing code or calling external interfaces.
Q: Is increased agent autonomy inherently risky?
A: Not necessarily. The research shows that most agent actions are low-risk, and agents often proactively seek clarification when uncertain.
Q: What industries are currently leading the adoption of AI agents?
A: Software development is currently the dominant sector, but other industries like finance and healthcare are rapidly adopting agent-based solutions.
Q: What is the key takeaway from this research?
A: Autonomy isn’t just about the model; it’s about the interaction between the model, the user, and the product.
Did you know? Agents are now requesting clarification more often than humans are interrupting them, suggesting a built-in safety mechanism.
Pro Tip: Focus on building robust monitoring infrastructure to understand how AI agents are being used in your organization.
Aim for to learn more about the future of AI? Explore our other articles or subscribe to our newsletter for the latest insights.
Keep reading