Three Agents at Once Locked My App's AI Out for Four Hours

Early one afternoon, the AI in my desktop app went quiet, and nothing in the app was broken. It stopped at about 13:20 on 19 October and stayed down until 17:10.

In a coding session I had three Opus agents building and reviewing at once, and between them they had spent the account limit the app runs on.

Everything shares one limit

My app's AI runs through the Claude Code CLI, on my own login. So does every chat session I open, every multi-agent workflow and every scripted sweep. They all draw on one account limit, nothing warns before it runs out, and the app's jobs just start failing.

It wasn't the first time, or the last:

• 14 October: two research workflows, about 8 million tokens across 41 agents, with a 212-call sweep running beside them. The app's AI was refused for about two and a half hours, until the reset.
• 19 October: three agents at once. About four hours.
• 20 October, about 23:00: only two agents, on the smaller Sonnet model, plus my own session. About 1.1 million tokens in 40 minutes, and four of the six agents still queued died at start.
• 24 to 27 October: the weekly limit this time. The app's AI was dark for about 90 hours, three mornings without a brief.
• 6 November, 01:30: my content system's nightly run found the limit already spent by the evening's agents. It reset at 02:00, half an hour later. Nothing waited for it, and the night made nothing.

The app was the smallest user

An app that runs AI jobs on its own all day looks like the big spender. An audit on 30 October said otherwise. Over the 15 days before it, the sessions building my command center used about 40 times what the app's own AI used. The weekly limit behind those 90 dark hours was spent by sessions, not by the app.

The same audit showed why the sessions cost so much. The median call in them re-read about 221,000 tokens of context, and one call in ten re-read over 720,000. So the fixes went into how I work with agents, and one went into the nightly run.

How I run several agents now

• At most two agents at a time, never three.
• No long sweep while agents run. A sweep runs on its own, sized to the task.
• Bulk work goes on a smaller model, and the mechanical part goes to a script. The morning after 20 October, a replacement table with tests did the mechanical part of a cleanup, and agents only did the judgment: a review by two skeptics, each with a short list of files.
• Sweeps can resume. After a reset, one call goes first as a probe, and the rest follow only if it answers.
• The agents I work with tell me when my app's AI will be blocked, and until when.

The night that waits for the reset

The limit message says when it resets (You've hit your session limit · resets 6:10pm), and nothing of mine read it. My content system's nightly run does now. It reads its batch report for the limit message, and if a 5-hour limit resets before 05:00, it waits for the reset and runs once more with what's left of its budget. A weekly limit or a reset in my morning gets no retry, only a note in the digest.

If you run agents next to an AI product, check which account each one draws on. Mine shared one, and the product was the one that went dark.

Four of the 25 checks in my runbook for AI that runs on its own come from these nights.