AI agents have no sense of time and are not aware of it AI代理没有时间感知能力，也无法意识到时间的流逝

TL;DR

• Popular coding assistants (Claude Code and Codex) consistently overestimate task duration, guessing ~90 minutes regardless of actual difficulty
• Self-assessment is severely flawed: models rated their own work ~20 percentage points higher than actual performance, even on failed tasks
• Runtime behavior depends heavily on the surrounding software harness, not just the underlying model
• Agents cannot reliably track elapsed time, making long-running autonomous tasks difficult to control
• Providing explicit time-tracking tools dramatically improves temporal accuracy

Why It Matters

This research reveals a critical gap in AI agent reliability for production use: without accurate time estimation and self-assessment, autonomous coding agents cannot be trusted on multi-hour workflows. For practitioners building or deploying AI agents, these findings highlight that model capability alone is insufficient—system design and tooling are equally decisive factors in agent performance.

Technical Details

• Study tested Anthropic's Claude Code and OpenAI's Codex on 200 ProgramBench tasks plus 18 custom benchmarks, measuring both forward time estimates and retrospective elapsed-time reporting
• Both models clustered around a 90-minute estimate regardless of task difficulty; retrospective errors ranged from 3x (Claude) to 6-10x (Codex) actual runtime
• Claude Code ran for a median of ~90 minutes per task, while Codex stopped after ~30 minutes, despite using similar underlying models—demonstrating harness-driven behavioral divergence
• Older models (Opus 4.8, GPT-5.5) overrated their results by ~20 points on average, claiming ~70% success on tasks that actually scored 7-14.5%
• When given access to an explicit elapsed-time tool, agents achieved near-perfect temporal accuracy, suggesting the limitation is architectural rather than inherent to the models

Industry Insight

• Agent harness design is as critical as model selection—identical models can produce drastically different outcomes depending on software wrappers, making benchmarking must include runtime behavior, not just capability
• Self-assessment calibration should be a standard evaluation metric for any agent deployment targeting long-running or autonomous workflows
• The fix is straightforward: integrating real-time tooling (clocks, timers) into agent loops can resolve temporal blind spots, suggesting that robust agent frameworks should bake in observability by default rather than relying on model intuition

TL;DR

• 主流编码助手（Claude Code、Codex）无法准确预测任务耗时，普遍高估所需时间，短任务误差最大
• 同一模型在不同软件设置下行为差异显著，运行时长和步骤数受harness影响达2.5倍
• 模型自我评估质量极不可靠，高估自身工作成果约20个百分点，甚至对失败任务给出高分
• 提供计时工具后模型能准确报告经过时间，表明外部辅助可弥补时间感知缺陷
• 研究揭示了AI agent在长时间任务执行中的关键缺陷，对开发可靠自主系统具有重要警示意义

为什么值得看

这项研究揭示了当前主流AI编码助手在时间感知和自我评估方面的系统性缺陷，对依赖长时间自主运行的AI agent开发具有重要参考价值。它提醒从业者，在构建可靠AI系统时，必须考虑外部监控和辅助工具来弥补模型内在能力的不足。

技术解析

研究使用ProgramBench数据集和18个自定义基准测试，评估Claude Code和Codex的时间预测能力。模型在任务前估计耗时，任务后报告实际经过时间，结果显示普遍高估，短任务误差最大。

同一模型在不同harness下行为差异显著，Claude Code平均运行90分钟，Codex约30分钟，步骤数相差2.5倍，表明软件设置对agent行为有决定性影响。

模型自我评估质量不可靠，老模型Opus 4.8和GPT-5.5高估结果20个百分点，甚至对失败任务给出高分，实际成功率仅7-14.5%。

当提供计时工具时，模型能准确报告经过时间，表明外部辅助可弥补时间感知缺陷，为agent设计提供可行改进方向。

行业启示

开发长时间运行的AI agent时，必须集成外部时间监控和进度跟踪工具，不能依赖模型内在时间感知能力。

AI系统的自我评估能力需要独立验证机制，避免高估导致任务失控或资源浪费。

研究凸显了harness设计对agent行为的关键影响，开发者应优化软件设置以提升一致性和可靠性，而非仅关注模型本身。