Where the Claude accounts went

What used up the shared Claude accounts this week, and the fixes ranked by how much they would save.

Oct 2 – 9, 2026 · 16 of 18 accounts at their weekly limit on Oct 9 · nothing changed yet

The biggest leak is five background "driver" sessions that have run non-stop for a week, plus the helpers they start. They are more than half of all usage.

Share of usage, by source

Background driver sessions and their helpers5 sessions running all week, thousands of helper sessions

57%

Your own sessions and their helpersthe ones you start and talk to

27%

GitHub PR reviewerabout 5,700 reviews, one on every push

10%

Agent SmithSlack replies and alert triage

Decision toolmany small sessions, little usage each

Suggested fixes, biggest first

#	Fix	What it means
1	Give background sessions a lifespan biggest win	Restart drivers daily, or once their history gets big, starting from a short summary instead of a week of history.
2	Start helper sessions fresh	Give each helper a short brief instead of a copy of its parent's whole history.
3	Let a small check answer automatic pokes	Only wake a big session when there is real work for it.
4	Review PRs once, not on every push	Review when a PR is marked ready. Also check the 714 reviews on Fede's branches, which came out of these accounts.

Why your teammate gets further, and how this was measured

Every reply re-reads a huge history. The heavy sessions re-read about 450,000 tokens (a few thousand pages) on every reply. A normal session re-reads a small fraction of that.
Almost nobody is typing. Only about 5% of usage came from a person. The rest is automatic pokes: PR events, nudges, "your screen hasn't changed" checks. Each poke pays to re-read the whole history.
Long gaps cost extra. A session woken after more than an hour has to rebuild its saved memory, which costs more. That rebuilding is about a third of all usage.
When one account runs out, the session moves to the next. So every account drains together instead of one.

Measured from the saved session logs on the Mac mini, GitHub's run history and the account switcher's log. Percentages compare usage at list price: they show relative size, not the actual bill. Which account each session used is only partly known, because sessions get moved between accounts.