Popular coding agents cannot reliably estimate how long a task will take or judge how well they completed it, according to a study of Anthropic’s Claude Code and OpenAI’s Codex. That uncertainty matters when developers hand agents long-running work and return only after it is supposed to be finished.
Two independent researchers tested the tools on 200 ProgramBench coding tasks and a separate set of 18 benchmarks. Before working, each agent predicted the required time; afterward, it estimated how long it had spent. Both often guessed about 90 minutes regardless of difficulty. In the second test, Claude’s estimates were off by about three times on average, while Codex missed by six to ten times.
The surrounding agent software, or harness, also changed behavior. Claude Code worked for a median of roughly 90 minutes, while Codex tended to stop after about half an hour. The same language model took 2.5 times more steps in Claude Code than in Codex.
Self-assessment was similarly weak. Older tested models overrated their scores by about 20 percentage points on average and sometimes judged failed work as largely successful. Developers therefore need external checks, real clocks and measurable completion criteria rather than relying on an agent’s confidence or time forecast.