Claude Code and Codex Agents Can't Judge How Long Their Work Takes

AI coding agents built on Anthropic's and OpenAI's latest models can write and fix code autonomously for hours, but they have no idea how long the work will take or how well they actually did it. That's the finding of a study by two researchers from the MATS 10 research program, reported by the-decoder.com.

What the researchers tested

Ofengenden and Andriushchenko tested the agents in two scenarios. In ProgramBench, the models had to rebuild working programs such as ffmpeg, SQLite, and ripgrep from scratch, without internet access. In AgentTime, a custom set of 235 tasks drawn from 18 different benchmarks including GPQA-Diamond, DeepSWE, and OSWorld 2.0, the agents tackled a mix of coding and research problems.

Before each task, the agent had to estimate how long it would take to complete. After finishing, it also had to judge how well it had done. The results were compared against the actual session duration and an objective assessment of solution quality.

Always around 90 minutes

Regardless of whether a task required a few minutes or several hours of work, the models consistently estimated it would take around 90 minutes. The error was largest for short tasks, and only for multi-hour tasks did the forecasts start to approach reality.

According to the-decoder.com, Claude was off by roughly a factor of three on average, while Codex was off by six to ten times. How time was tracked also turned out to depend on the tool itself, not just the underlying language model: Claude Code worked for a median of around 90 minutes in practice, while Codex stopped working after about 30 minutes regardless of task complexity.

A false sense of success

An even clearer problem showed up in the agents' self-assessment of work quality. The older models, Opus 4.8 and GPT-5.5, systematically overstated their scores by around 20 percentage points relative to the actual quality of the solution. In one case described in the study, both models rated their own work at around 70 percent success, while the actual results came in at 7 and 14.5 percent.

Both models give themselves high ratings even on tasks that ended in failure - from the study by Ofengenden and Andriushchenko

Newer model versions performed differently. Opus 5 tended to underestimate its own results by 11 to 15 points, while GPT-5.6 Sol overstated its score by only about 7 points, considerably less than its predecessor. This suggests that self-assessment calibration improves with each new model generation, though it is still far from precise.

A clock as a substitute for a missing sense

The authors also checked what happens when an agent has access to a tool that shows the actual elapsed time while it works. In that case, the accuracy of its estimates rose sharply, the model almost always correctly judged how much time it had spent on a task. Without access to a clock, the agents relied mainly on the length of the generated conversation transcript, the correlation between transcript length and actual session duration reached 0.91.

The researchers also introduced a metric they call the compression exponent, describing how far a model's predictions deviate from the actual time scale of tasks. For Claude it came out to 0.24, and for Codex 0.19, meaning that in both cases the models treated short and long tasks in a very similar way, instead of scaling their predictions proportionally to task difficulty.

Why this matters for companies deploying agents

The results carry practical weight for companies that increasingly let AI agents work autonomously for extended periods, setting time or budget limits per task. If a model doesn't know how long it needs and also overrates the quality of its own work, it becomes hard to build reliable mechanisms for tracking progress or automatically halting failed attempts.

The problem also affects Polish teams deploying coding agents in day-to-day work, where decisions about when to stop an agent's session and hand the task back to a human often rely on the model's own claims. The study suggests that a simple and effective fix is equipping the agent with an external time-tracking tool, rather than relying on its own judgment.

The authors stress that the lack of a sense of time doesn't stem from limits in the models' reasoning itself, but from the absence of a signal that humans have naturally. Agents don't feel fatigue, don't see a wall clock, and have no built-in circadian rhythm, so without explicitly provided time information they have to guess based on patterns from their training data.