AI agents have no sense of time and are not aware of it
decoder ·
A new study by independent AI researchers, conducted under the MATS program, found that coding assistants like Anthropic's Claude Code and OpenAI's Codex have a poor sense of time. Before tasks, they consistently overestimated how long they would need, often guessing around 90 minutes regardless of the task's actual difficulty. In tests, Claude's estimates were off by an average of three times and Codex's by six to ten times, with the worst inaccuracies on short tasks. The agents' runtime also varied greatly depending on the software harness used, with Codex stopping after about 30 minutes and Claude Code working up to 90 minutes. Additionally, the agents overrated their own performance, giving themselves scores about 20 percentage points higher than their actual results. The researchers stress that accurate self-assessment is crucial for controlling agents on long-running tasks, and note that when given a tool that reports elapsed time, the agents estimated it correctly almost every time.
decoder