AI Agents Found Lacking Time Perception

AI coding assistance tools lack a "sense of time," and moreover, they don't realize it themselves——such research findings have emerged. The subjects of investigation were Anthropic's Claude Code and OpenAI's Codex. Both were confirmed to have a tendency to systematically overestimate the time required for tasks.

An AI agent is an AI system that executes multiple steps autonomously toward a given goal without humans providing detailed instructions. In the case of coding assistance tools, it performs a series of operations—code generation, modification, and testing—automatically. As such autonomous task processing increases, whether AI can accurately grasp "how much time will be required" is becoming an important practical issue.

What stood out particularly in this research was the magnitude of Codex's error margin. Cases were found where it estimated times up to 10 times longer than the actual duration required. Claude Code showed a similar trend, suggesting a structural problem common to both. What is even more concerning is the accuracy of self-assessment. Both tools tend to evaluate their own work quality approximately 20 points higher than reality.

This "overoptimistic self-assessment" goes beyond being merely a performance problem. In scenarios where AI agents proceed autonomously over extended periods, AI's self-report is sometimes used as a basis for determining when humans should intervene. If an AI incorrectly reports "I am performing well," there is a risk that problems will be discovered late or opportunities for correction will be missed.

It is important to note the current rapid expansion in the use of such AI agents. Beyond simple question-and-answer interactions, use cases are increasing where humans delegate tasks requiring long hours of work—such as automated code generation, data analysis, and workflow automation. Correspondingly, demands are rising for AI to accurately understand its own status and work progress.

This research suggests that current AI agents may not yet be adequately equipped with "metacognition"——the ability to objectively understand their own capabilities and limitations. Inaccurate time estimation and self-assessment will have implications for how monitoring and management systems for AI are designed. As AI agent practical applications advance in the future, how to improve the accuracy of such self-awareness will become a notable focus for developers and adopting organizations.

From the user's perspective, rather than delegating tasks entirely to AI agents, incorporating periodic progress checks and reviews of deliverables appears to be a more realistic approach for now. It requires maintaining a critical stance that does not simply accept AI's "I have completed this" reports at face value.