Claude Code subagent hit 約87万トークン; split by task

The subagent structure exists to keep each delegated task's context separate, but the author's logs show this only pays off when the split follows the kind of work. In the largest session, about 5.6 million tokens of final Context spread across 27 subagents dwarfed the parent's roughly 400,000, and the outliers were an on-device check and test-following work.

The three fixes point in different directions. One kept test writing with the implementation agent; another split test clean-up away from it. The author explains both by asking where the bloat sits: reduce the number of on-device checks in one case, and stop several task types from piling up in one agent in the other. The rule of thumb that emerged was to size a single delegation at one document or one change.

Whether these gains hold is not settled. The sessions after the switch to Opus 5.5 had different conditions and were not a clean model comparison, and the author notes the cost of over-splitting is more request and report overhead. The figures here are one developer's own measurements, useful as a way to see where usage goes rather than as a benchmark for other setups.