AI Agents Have No Sense of Time — and Have No Idea That's a Problem
Agentic AITell an AI agent to work on a task for two hours. It will not work for two hours. It doesn’t know what two hours is.
New research from MATS program researchers, covered by The Decoder, tested Claude Code and OpenAI’s Codex on their temporal awareness — both before and after task execution. The results reveal a gap that grows more significant as agentic deployment scales.
Why Everyone Assumes Agents Understand Time
The assumption is intuitive: if an AI can reason about code, schedules, and deadlines, it surely has some model of time passing. Agents write commit messages with timestamps, follow multi-step plans, and report task completion. From the outside, they look temporally aware. The problem is that apparent awareness isn’t the same as actual perception of elapsed time.
Large language models don’t have internal clocks. There’s no mechanism for a model to sense that 47 minutes have passed since it began a task — not because the engineers forgot to add one, but because inference is stateless. Each forward pass has no inherent connection to wall-clock time.
What the Evidence Shows
The researchers tested both models on 200 tasks from ProgramBench plus 18 custom assessments — 218 total benchmarks. The pattern: “On ProgramBench, both models mostly guessed around 90 minutes, no matter the difficulty.” Short tasks and long tasks produced nearly the same estimate because there was no actual measurement happening.
When researchers checked elapsed time retrospectively, Claude Code underestimated by roughly 300%. Codex was off by 600 to 1000%. The infrastructure hosting the model mattered significantly — Claude Code maintained median runtimes around 90 minutes, while Codex consistently stopped after approximately 30 minutes regardless of complexity. The same underlying models produced dramatically different behaviors depending on their harness.
Self-assessment compounded the problem. Older models (Opus 4.8 and GPT-5.5) overrated their own results by 20 points on average. In one striking case, both systems estimated their success rate at approximately 70%, when actual performance measured 7% and 14.5% respectively.
The Correction
The fix turned out to be simple: give agents access to a tool that reports elapsed time, and “they got it right almost every time.” The problem isn’t architectural — it’s that temporal awareness requires an explicit mechanism. Without it, agents invent an estimate from priors (tasks feel like 90 minutes) rather than measuring.
This has a practical implication for anyone building agentic pipelines: time constraints and duration targets need to be enforced by the harness, not trusted to the agent’s self-report. Telling an agent to “finish this in 20 minutes” without a mechanism to enforce or even measure that constraint produces no reliable result. The agent will complete what it completes, then report that it took whatever it estimates was 20 minutes.
The researchers are now investigating whether agents can maintain adherence to predetermined work durations when explicitly programmed for that constraint — which is a different question than whether they can perceive time at all.
The So What
Two changes that matter for production agentic deployments: first, never rely on an agent to self-terminate based on a time budget without external enforcement. Second, treat agent self-assessment of success with the same skepticism you’d apply to time estimates — the 7% vs 70% gap in this research is a reminder that confidence and accuracy have almost no correlation in agentic task completion. Build your own evaluation hooks. Don’t ask the agent whether it succeeded.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →