Blackbox Blast
BB Preflight
← Incident log

CASE · Minor · runaway-loop

Unverified reportCollected from public reports

/code-review fanned out to 10 subagents and burned a 5-hour quota in 4 minutes (unverified report)

An automatic pre-push /code-review in Claude Code started 10 subagents at once instead of one. The first run used up the 5-hour session limit in 8 minutes, with six of the ten cut off; the second did it in 4, with all ten cut off.

3 min read

OBSERVED
CATEGORY
runaway-loop
BLAST RADIUS
financial, uptime
DETECTED BY
quota-anomaly
TAGS
#claude-code #subagents #code-review #usage-limit #model-upgrade #parallelism

CH 1

Cause

Per the author, three changes landed on the same day: the model moved from Opus 5 to Opus 5.5, effort was raised from high to max, and the desktop app updated (2.1.275 to 2.1.280). Better delegation plus a wider review scope, under a default ceiling of 20 concurrent subagents, produced the fan-out. The exact trigger was never isolated.

CH 2

Consequence

26 subagents across two incidents, 16 of them cut off before finishing, and the weekly limit hit on the third day after the new model launched (the author cannot say the ten agents alone caused it). The app estimated the whole conversation at about $31 at API prices, not actually paid on the flat-rate Pro plan; subagents made up about 40% of the conversation's usage.

CH 3

Fix

The author replaced the multi-agent review with a single subagent that checks correctness only; after the main session's own pass found one bug, it found two more real ones. Not yet tried: lowering the concurrent-subagent limit, turning off automatic invocation, lowering effort.

What happened

The author, a small-business owner who is not a programmer, had a rule: whenever the calculation logic of a home-made tool that produces figures for client documents changed, Claude Code would run /code-review on its own before pushing. On September 23, 2026, that review suddenly started ten subagents at the same time instead of the usual one. Eight minutes later the 5-hour session limit was gone.

The chaos on the ground

The next attempt was worse: the limit ran out in four minutes and all ten subagents stopped halfway, so the review produced nothing usable. Across the two incidents 26 subagents were started and 16 were cut off. On the third day after the new model came out, the weekly limit was hit as well.

The author checked the app's usage breakdown and saw that subagents accounted for 40% of that conversation. Asked to count its own logs, Claude found that the 23 reviews from July to September 17 had used one subagent at most; ten had never been started before September 23.

Root cause

Three things changed at once that day: the model (Opus 5 to Opus 5.5), the effort setting (high to max), and the app version. The author's reading is that the newer model delegates more readily, max effort widens what a review covers, and nothing stopped it from running many agents at once, since the default limit allows 20. Because all three changed together, there is no way to say which one tipped it over.

The fix

The author dropped the multi-agent review and wrote a rule into Claude's memory: pre-push reviews go to a single subagent with no prior context, limited to correctness: reading line by line, behaviour changes, tracing across files, language pitfalls and value handling.

That narrower review found two more genuine bugs after the main session's own pass had found one. Lowering the concurrency limit, turning off the automatic review, or lowering effort were left untried.

For anyone automating agents, two habits would have made this cheaper: change one of model, effort and version at a time so the cause of a jump is visible, and set a low ceiling on parallel agents before a smarter model decides to use all the room it has.

Sources

PREFLIGHTIs your own Mac prepared?BB Preflight is a free Mac tool that checks 21 things, from AI agent permissions to where your keys live, without changing anything.Run the check →