Coding agents misjudge task duration and overrate their own work
A new study reported by The Decoder finds that widely used AI coding assistants such as Claude Code and Codex lack a reliable sense of time when estimating how long tasks will take. The research describes systematic duration misjudgment: Codex predictions can diverge from actual runtimes by as much as tenfold, while agents also tend to score their own output roughly twenty percentage points higher than warranted. For builders running longer autonomous coding sessions, that combination weakens planning, scheduling, and human oversight, because completion signals and self-assessment may look credible even when timing and quality judgments are badly miscalibrated. The coverage summarizes attributed findings rather than releasing primary data here, but the reported gap is large enough to matter for anyone treating agent time estimates or self-ratings as operational ground truth.
Coding agents misjudge task duration and overrate their own work
AI coding assistants like Claude Code and Codex have no sense of time, according to a new study. Codex is off by as much as ten times the actual duration, and they also rate their own work about 20 percentage points too high.
Key takeaway
Coding agents can misread both how long work will take and how good it is, so autonomous runs need external checks instead of trusting agent timing or self-ratings.
What happened
According to reporting in The Decoder, a new study concludes that AI coding assistants including Claude Code and Codex have no sense of time and are not aware of it when judging task duration.
The Decoder says both tools systematically overestimate how long tasks will take, with Codex off by as much as ten times the actual duration, and they rate their own work about 20 percentage points too high.
Evidence
AI coding assistants like Claude Code and Codex have no sense of time, according to a new study.
The Decoder · attributed
AI coding assistants like Claude Code and Codex have no sense of time, according to a new study.
Both systematically overestimate how long tasks will take.
The Decoder · attributed
Both systematically overestimate how long tasks will take.
Codex is off by as much as ten times the actual duration.
The Decoder · attributed
Codex is off by as much as ten times the actual duration.
They also rate their own work about 20 percentage points too high.
The Decoder · attributed
They also rate their own work about 20 percentage points too high.
For long, autonomous tasks, that creates a real problem with oversight.
The Decoder · attributed
For long, autonomous tasks, that creates a real problem with oversigh
Why it matters
Long autonomous coding workflows depend on credible time estimates and quality signals; large timing and self-rating errors can leave operators underprepared to intervene.
Limits and uncertainties
The packet provides only The Decoder's summary of a new study, not the underlying paper, methods, sample size, or benchmark details.
The excerpt on oversight is truncated in the source material, so the full scope of the oversight concern is not fully visible here.
Practical implications
Do not use agent duration forecasts as scheduling ground truth without independent timing measurements.
Treat agent self-assessments of code quality as weak signals and keep human or automated review gates for longer autonomous runs.
What to watch
Publication of the primary study with benchmarks, models tested, and methodology behind the tenfold duration error and 20-point self-rating gap.
