AI agents are increasingly asked to handle coding work that can run beyond a quick prompt-and-answer exchange. A study described by THE DECODER shows a basic problem with that shift: popular coding assistants can be poor judges of time, both before and after they work.
The study looked at Anthropic's Claude Code and OpenAI's Codex. The finding was not just that their forecasts were inaccurate. The larger issue is that the agents did not reliably know how long a job would take, and they also struggled to report how long they had already spent.
What the study tested
The research was conducted by two independent AI researchers as part of the MATS research program. They tested the time awareness of two widely used coding assistants, Anthropic's Claude Code and OpenAI's Codex.
The setup was direct. Before each coding task, the AI agents had to estimate how much time they would need. After completing the task, they had to look back and report how much time had passed.
The test set included 200 tasks from ProgramBench, along with the researchers' own suite of 18 benchmarks. That gave the researchers two ways to examine the same broad question: whether coding agents can reason about time in a way that makes them easier to manage during long-running jobs.
The answer, based on the reported results, was mostly no. The agents tended to overestimate the time they would need. On ProgramBench, both models often gave estimates around 90 minutes, even when task difficulty varied.
Where the time estimates failed
The errors were not small calibration misses. In the second round of tests, Claude was off by three times on average. Codex was off by six to ten times.
The worst estimates appeared on short tasks. Only when tasks reached the multi-hour range did some predictions move closer to reality. That pattern matters because many real workflows include a mix of quick fixes, medium tasks, and longer jobs. If an AI assistant treats many of them as roughly the same length, planning becomes harder.
For a user, a bad time estimate is more than a cosmetic flaw. Time is part of how work is scoped, reviewed, interrupted, and trusted. If an agent says a task will take much longer than it does, the user may plan around a false delay. If it cannot accurately report elapsed time after the fact, the user has less visibility into what actually happened.
The study also points to a deeper control problem. A long-running coding agent is not just producing code. It is deciding when to continue, when to stop, and how much effort to spend. Weak time awareness makes those decisions less dependable.
The harness changed the behavior
The study found that results shifted depending on the software setup around the model. Claude Code and Codex did not behave the same way when running through their respective systems.
Claude Code kept working until it judged the task complete, with a median of about 90 minutes. Codex stopped after roughly half an hour, almost regardless of the task.
The surrounding software is known as the harness. According to the study, the same language model takes 2.5 times more steps in Claude Code than in Codex on average. That means runtime is shaped by both the model and the system that manages it.
This is important for anyone evaluating AI coding assistants. A model's behavior cannot be separated cleanly from the workflow it runs inside. The same underlying language model may act differently depending on the rules, stopping conditions, tools, and execution loop around it.
Self-assessment was also weak
The time problem was not isolated. The agents were also unreliable when judging the quality of their own work.
The older models, Opus 4.8 and GPT-5.5, overrated their results by 20 points on average. They also gave themselves high marks even when tasks failed.
In one case, both models estimated that their work was about 70 percent successful. The actual scores were 7 and 14.5 percent.
That gap matters because self-assessment is part of autonomy. An agent that cannot evaluate its own progress may continue when it should revise, stop when it should continue, or report confidence that the result does not deserve.
For coding work, the practical risk is straightforward. If an assistant misreads both time and quality, a user may need stronger external checks: tests, review steps, time tracking, and clearer stop conditions.
Why external time tools matter
The researchers say self-assessment is important for agents that are expected to work reliably on long tasks. One example instruction is: "iterate on this task for two hours." That type of request depends on an agent understanding elapsed time well enough to obey the limit.
An agent that misjudges time is harder to control. It may not know whether it has spent minutes or much longer. It may also fail to connect the remaining time budget with the quality of the work still needed.
The study also showed a simple path to improvement. When the agents were given access to a tool that reports elapsed time, they got it right almost every time.
That finding suggests the problem is not necessarily solved by asking agents to reason harder about time internally. A more practical approach is to give them a reliable clock and make that clock part of the workflow.
The researchers want to test next whether agents can stick to a set work duration. That question follows naturally from the results. Knowing elapsed time is one thing. Using it to follow a work schedule is another.
For now, the takeaway is clear: AI agents can be useful coding assistants, but long-running work needs guardrails. Accurate timing, external tools, and independent quality checks are still central to making AI agent workflows manageable.