New Oxford Research Finds LLM Agents Largely Lack Time Awareness

Oxford researcher Maksym Andriushchenko published a blog study "Are LLM agents time-aware?" on August 18, systematically examining LLM agents' temporal awareness. The core finding: frontier models have almost no concept of time—they can neither predict how long tasks will take in wall-clock terms nor estimate how long they have already spent, making lack of time awareness a notable failure mode in long-horizon tasks.

Confirmed

• The study centers on two questions: whether agents can predict wall-clock time of tasks, and whether they can estimate their own elapsed time
• Experiments cover long-horizon benchmarks including ProgramBench, PaperBench, and DeepSWE with clean experimental setups
• When estimating task completion time, Claude Code and Codex estimate "human expert" time as 3 to 4 times higher than their own, i.e., models tend to believe humans are 3–4x slower than themselves
• The author argues poor time estimation stems from agents not experiencing time as humans do; commenter Ziv Ravid speculates limitations in numeric handling and the inherent difficulty of prediction may also contribute
• The author shares a habit of regularly testing models' time understanding, including concrete results for the new Qwen3.8 27B model
• The study cites a LessWrong post on digital workers' time perception and situational awareness

Unconfirmed

• One repost claims the study found LLMs "already have time awareness," contradicting the original conclusion that frontier models lack a concept of time; the original blog should be treated as authoritative

Why it matters

• Time planning and estimation underpin autonomous long-horizon task execution and sensible allocation of compute and action budgets; the study offers a new empirical lens for evaluating and improving agent self-awareness
• The Qwen3.8 27B test shows this evaluation method can directly track time-awareness of new models

2026-08-18 ~ 2026-08-18 · 9 related posts

• [source] New research: are LLM agents time-aware? Can they estimate task duration? — maksym_andr · 2026-08-18
• Research: Are LLM agents time-aware and can they estimate task duration? — xeophon · 2026-08-18
• [source] Agents struggle with time estimation: testing Qwen3.8 — maksym_andr · 2026-08-18
• Study Finds LLM Agents Lack Time Awareness Near Deadlines — maksym_andr · 2026-08-18
• LLMs estimate human task time 3-4x longer than their own, study finds — maksym_andr · 2026-08-18
• [source] Frontier models lack time awareness, failing to estimate task duration — maksym_andr · 2026-08-18
• New Research: LLM Agents Lack Time Awareness, Can't Estimate Hours Spent on Tasks — solyarisoftware · 2026-08-18
• Study: LLM Agents Lack Time Awareness for Task Duration — ziv_ravid · 2026-08-18

1 near-duplicate retellings: maksym_andr