Skip to content
Ben Dangelmayr

Case study: Profiling coding-agent runs

This page counts all 2,977 subagent runs from June 15 to October 27, 2026. 2,818 of them are the runs of the follow-up measurement in the case study (cutoff October 26, 12:00 UTC); the other 159 started later. This page counts wait loops per 100 shell commands, which are 44 percent of all tool calls; the case study counts per 100 tool calls of any kind, so its rates are about half as high (1.16 until August 7, 0.82 in October). It also compares calendar periods of direct subagent runs instead of the periods between the rule changes. The model share here is measured against the total time of all runs, there against eight selected runs.