Nice work! I like the ablations, and the harness-level comparison is an important investigation. However, the post presents its main findings without considering very similar results from other papers. A lot in your analysis is genuinely new, but the headline has been shown before.
Can LLMs Perceive Time? An Empirical Investigation (ICLR 2026 workshop) tests wall-clock prediction and finds LLMs overestimate their own runtime by ~4–7x, similar to the 3x/7x you report. They use a ReAct scaffold though, instead of Claude Code or Codex.
Your LLM Agents are Temporally Blind (Findings of ACL 2026) tests LLMs’ alignment with human’s judgments in time-sensitivity scenarios. Like you, they test adding timestamps and also find that alignment rises, though not dramatically.
Discrete Minds in a Continuous World (Findings of EMNLP 2025) tests where a sense of time could come from and finds that models can map conversation token counts to elapsed wall-clock time (their “token-time hypothesis”, cf. your transcript-length correlation of r = 0.91), and adapt behaviour under expressed urgency (loosely similar to your human/AI ablations).
Hi, thank you for your comment. Yes, we’ve actually read all three of these papers and more! This blog post serves to see if these conclusions still stand across new frontier coding agents on very long horizon tasks, Can LLMs Perceive Time evaluates on tasks that often take less than a few minutes. Our graph shows that the estimation of tasks depends on task length. Your LLM Agents are Temporally Blind almost purely depend on whether agents receive time-stamps in their system prompts. We want to examine behavior on long-horizon tasks where agents have all the tools to their disposal.
An even more recent paper, BAGEN: Are LLM Agents Budget-Aware? came up with opposite conclusions for budget estimation. We would like to see if ours better reflect experiences that users have with coding agents. We also find these papers don’t include the harnesses and ablations that we want to test against (we’ll expand to Opencode and Chinese ones). Finally, we’re also runing an updated wall-clock duration-following benchmark, which is emphasized in next steps. We will also make sure to include a very thorough related works section in our coming paper!
Nice work! I like the ablations, and the harness-level comparison is an important investigation. However, the post presents its main findings without considering very similar results from other papers. A lot in your analysis is genuinely new, but the headline has been shown before.
Can LLMs Perceive Time? An Empirical Investigation (ICLR 2026 workshop) tests wall-clock prediction and finds LLMs overestimate their own runtime by ~4–7x, similar to the 3x/7x you report. They use a ReAct scaffold though, instead of Claude Code or Codex.
Your LLM Agents are Temporally Blind (Findings of ACL 2026) tests LLMs’ alignment with human’s judgments in time-sensitivity scenarios. Like you, they test adding timestamps and also find that alignment rises, though not dramatically.
Discrete Minds in a Continuous World (Findings of EMNLP 2025) tests where a sense of time could come from and finds that models can map conversation token counts to elapsed wall-clock time (their “token-time hypothesis”, cf. your transcript-length correlation of r = 0.91), and adapt behaviour under expressed urgency (loosely similar to your human/AI ablations).
Hi, thank you for your comment.
Yes, we’ve actually read all three of these papers and more! This blog post serves to see if these conclusions still stand across new frontier coding agents on very long horizon tasks, Can LLMs Perceive Time evaluates on tasks that often take less than a few minutes. Our graph shows that the estimation of tasks depends on task length. Your LLM Agents are Temporally Blind almost purely depend on whether agents receive time-stamps in their system prompts. We want to examine behavior on long-horizon tasks where agents have all the tools to their disposal.
An even more recent paper, BAGEN: Are LLM Agents Budget-Aware? came up with opposite conclusions for budget estimation. We would like to see if ours better reflect experiences that users have with coding agents. We also find these papers don’t include the harnesses and ablations that we want to test against (we’ll expand to Opencode and Chinese ones). Finally, we’re also runing an updated wall-clock duration-following benchmark, which is emphasized in next steps.
We will also make sure to include a very thorough related works section in our coming paper!