AI Agents Have No Sense of Time — And They Don't Know It

Most teams deploying AI coding agents assume they can give a time constraint — “iterate on this for two hours” — and get predictable behavior. The assumption is wrong, and the agents don’t know it’s wrong.

Why Everyone Believes Agents Can Follow Time Constraints

The belief is reasonable. Agents respond to natural language instructions. Developers give human teammates time-bounded tasks constantly — “spend 30 minutes on this,” “we have until end of sprint” — and humans generally comply. The intuition transfers because agent interfaces look like human communication surfaces. They are not.

AI coding agents lack a functional internal clock. They don’t know how much time has elapsed during a task. They don’t know how much time remains. They cannot reliably estimate either, even when asked directly.

What the Evidence Shows

Research on coding agents using ProgramBench tasks produced findings that should recalibrate anyone running agents in time-sensitive production workflows.

Both Claude Code and Codex “consistently overestimated how much time they’d need” — and then diverged in behavior. Codex stops after approximately 30 minutes almost regardless of task complexity. Claude Code continues for a median of around 90 minutes. Both systems guessed approximately 90 minutes as their expected duration regardless of actual task difficulty, which means the estimate is essentially static rather than calibrated.

The self-assessment failure is more striking. In one measured case, both systems scored themselves at “70 percent successful” when actual task completion rates were 7% and 14.5% respectively. Agents operating without external time tools not only can’t track time — they significantly overestimate the quality of work they produce within the time they’ve used.

Older models showed similar patterns: “overrated their results by 20 points on average” across evaluations.

The proposed solution from the research is structurally simple: when agents had access to “a tool that reports elapsed time, they got it right almost every time.” External time-tracking mechanisms, not model-side improvements, appear to be the practical mitigation.

The Correction

AI agents cannot be trusted to self-regulate toward a time boundary without an external enforcement mechanism. “Iterate for two hours” is not a well-formed instruction for a production agent. The agent will continue until it decides to stop — which, for Claude Code, is typically around 90 minutes, and for Codex, typically around 30 — driven by internal heuristics that have nothing to do with the actual clock.

The corrected operational model: provide agents with an external time-reporting tool, set hard cutoffs at the infrastructure layer rather than the instruction layer, and treat any agent self-assessment of task completion as an indicator to verify rather than a signal to trust. The Decoder’s full analysis covers the underlying research in detail and includes the methodology on the ProgramBench evaluation set.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.