AI agents have no sense of time and are not aware of it

Details

Two independent AI researchers, working through the MATS research program, tested Claude Code and OpenAI's Codex on 200 tasks from ProgramBench plus 18 custom benchmarks. Neither agent could accurately estimate how long tasks would take: both defaulted to guessing around 90 minutes regardless of difficulty, with Claude's estimates being 3x off actual duration on average and Codex's being 6-10x off — with the worst errors on short tasks. The study identifies a second problem beyond bad time predictions: the models cannot reliably assess the quality of their own work. Older models (Opus 4.8 and GPT-5.5) graded their own output about 20 percentage points above actual scores. In one test case, both self-reported roughly 70% success when actual scores were 7% and 14.5% respectively. A third finding concerns harness effects: the same underlying model takes 2.5x more steps inside Claude Code than inside Codex. Runtime behavior depends as much on agent scaffolding as on the model itself — meaning benchmark results from one harness may not transfer to another. For practitioners building autonomous agents running multi-hour jobs, this matters: agents given instructions like 'work on this for two hours' cannot reliably comply when they cannot track elapsed time or judge whether they've succeeded. The researchers conclude this is a fundamental architectural gap, not a calibration issue that fine-tuning can easily fix.