Predicting Agentic Coding Task Duration from Prompt-Time Features: A Single-Developer Instrumented Study

Abstract

A single-developer, exploratory study testing whether the duration of an agentic coding turn can be predicted from features available at prompt submission (prompt length, task type, repository scope, session history), using 978 real turns logged over 12 days. Three trained models (elastic net, gradient-boosted trees, gradient-boosted quantile regression) were compared against a simple baseline that predicts a turn's duration from the previous turn in the same session. None of the trained models beat that baseline. Prediction intervals meant to cover 80% of outcomes only covered about 58%. No pre-registration was filed, so results are reported as exploratory findings, not confirmed hypotheses. Collector, analysis code, and dataset are released for replication.