AI agents are shifting work from chatbots to autonomous task execution, study finds

Frontier AI models from Anthropic, OpenAI, and Google are improving at a rate described as better than exponential, according to assessments from METR, the UK’s AI Security Institute, and GDPval. These evaluations measure how much human work AI can perform with a single prompt, with results showing rapid gains in capability. Epoch’s experiments found that Opus 4.7, operating autonomously for 14 hours, completed a software package estimated to require 2–17 weeks of human engineering effort, at a token cost of $251. While not universally reliable or cost-effective, the trend indicates accelerating progress in AI’s ability to execute complex, long-duration tasks.

The operationalization of AI is shifting from interactive chatbots to autonomous agents, which are equipped with tools, environments, and specialized applications like Claude Code or OpenAI’s Codex. This transition is documented in internal OpenAI data, where a quarter of employees report running at least four agents weekly. Notably, adoption is not limited to technical roles: legal, HR, and other non-technical functions are adopting agents at nearly the same rate as engineers, suggesting a broad-based shift in workflow integration.

Success with agentic tools appears tied to domain expertise rather than job title. A study of Claude Code users found that professionals with deeper domain knowledge achieved higher success rates and extracted more useful output per prompt, regardless of whether they were software engineers. This implies that the most effective use of AI agents comes from experts who can guide and refine AI-driven processes, effectively acting as managers of autonomous systems.

The rapid improvement in AI capabilities is creating a mismatch with institutional adaptation. Organizations accustomed to planning for incremental AI assistance are now facing scenarios where a single prompt can deliver 16 or more hours of autonomous work. This acceleration—felt as sudden leaps rather than smooth curves—is driving turbulence in policy, market expectations, and strategic planning, as institutions struggle to keep pace with capabilities that double in fixed windows.