Dynamic concurrency is the switch that lets a coding agent decide, mid-run, to fan work out to sub-agents. Three harnesses were run over the same 354 tasks twice each, once with that switch on and once off, with model, prompt, environment and time budget held constant. On bounded bug-fixing two of the three got significantly worse with it on — 59% of tasks passed against 83%, and 59% against 77% — while tokens ran 1.41 to 3.31× higher and only 27.9% of jointly-solved tasks finished faster in parallel. It earns its keep at the far end of the horizon, where one harness went from 7.1% to 21.4% on the most sustained benchmark. And of the 804 failure instances annotated across 1,062 concurrent traces, the largest single bucket is not a planning subtlety: 26.5% were sub-agents editing the same or related files with nobody coordinating them.