Study Finds AI Coding Agents Cannot Track Elapsed Time or Judge Their Own Work Quality

The Finding

AI coding agents have no reliable internal sense of how much time has passed while they work, according to a study by two independent researchers conducted through the MATS research program. The pair tested Anthropic's Claude Code and OpenAI's Codex and found both systems consistently and badly misjudge task duration — a problem that undermines efforts to run agents on long, multi-hour jobs.

How the Test Worked

Before each coding task, the agents were asked to estimate how long they would need to complete it. After finishing, they were asked to report how much time they believed had actually elapsed. The researchers drew test material from 200 tasks in a benchmark collection called ProgramBench, supplemented with their own suite of 18 benchmarks.

The results were consistently poor. On ProgramBench, both models guessed roughly 90 minutes for nearly every task, regardless of actual difficulty. When asked to look back and estimate elapsed time, Claude was off by a factor of about 3x on average, while Codex missed by 6x to 10x. Errors were worst on short tasks; only in the multi-hour range did some estimates come close to the real runtime.

Behavior Depends on the Harness, Not Just the Model

The study also found that runtime behavior is shaped more by the surrounding software environment — the "harness" — than by the underlying model itself. Claude Code kept working until it judged a task complete, with a median runtime of about 90 minutes. Codex, running the same underlying capability, stopped after roughly 30 minutes almost regardless of task complexity. According to the researchers, the same language model took 2.5 times more steps when run inside Claude Code compared to Codex.

Agents Also Overrate Their Own Work

Beyond time perception, the study found agents are equally unreliable at judging the quality of their own output. Older models tested — Opus 4.8 and GPT-5.5 — overrated their results by roughly 20 percentage points on average, according to the researchers, and in some cases assigned themselves high marks on tasks they had actually failed. In one instance, both models estimated their work was about 70 percent successful; the actual scores were 7 percent and 14.5 percent respectively.

A Fix: Give the Agent a Clock

When the researchers gave the agents access to a tool that directly reports elapsed time, accuracy improved dramatically — the agents got their time estimates right almost every time. This suggests the core issue isn't reasoning capability but a lack of grounding: without an explicit signal, agents have no mechanism to track duration at all.

The researchers say this matters for practical deployment. For an agent to reliably follow instructions like "work on this for two hours," it needs an accurate internal sense of how much time has passed. Without that, long-running autonomous tasks become hard to control or budget for. The team's next step is testing whether agents can be made to stick to a specified work duration when given proper time-tracking tools.

What This Means

This research exposes a gap between how AI agents are marketed — as capable of autonomous, hours-long work — and their actual internal state awareness. Agents don't perceive time passing any more than they perceive their own success; both are inferred, often badly, from training patterns rather than grounded observation. The practical fix is straightforward — expose elapsed time as an explicit tool call — but the underlying issue is broader: agents lack reliable self-monitoring unless developers explicitly engineer it in. Companies building long-running autonomous coding agents will need to treat time-awareness and self-evaluation as separate infrastructure problems, not capabilities that emerge naturally from bigger models.