THE AGENTS KEPT SPENDING.

We first logged the cluster on September 20, after three separate public Codex reports showed automated agent or subagent work chewing through large portions of weekly allowance — during a bounded task, or after the user believed work had stopped. Different failure shapes, so this is a pattern report, not one bug behind all three.

On September 16, one Codex Desktop user reported 1,545 model responses and 242,318,802 telemetry-reported tokens over 2 hours 44 minutes — while the account's weekly usage meter rose about 11 percentage points. The report says 98.1% of its measured input tokens were cached and explicitly warns that raw local telemetry should not be treated as a billing statement or a one-to-one measure of subscription usage. An earlier Codex issue had already described fixed context, tool and skill overhead multiplying under subagent fan-out.

A separate September 20 Codex Security report says a freshly reset weekly allowance went from 0% to 100% in about 44 minutes while discovery workers and subagents kept being created. The report says discovery-0013 was launched when weekly usage was already around 97–98%, and another worker was attempted after the weekly limit had been exhausted. OpenAI's current help documentation says Work and Codex can use five-hour and weekly allowances and that a reset refreshes the weekly allowance — but it does not describe a task-wide fan-out budget for this incident.

The third report is a different stop-boundary failure. After a natural-language request to pause, Codex acknowledged the pause — but the active Goal generated 2,466 continuation turns over roughly five more hours; 2,461 of the final responses were empty. Account-wide weekly usage readings rose from 58% to 98% during that period, but the reporter says those readings cannot prove the single task caused the entire increase and notes that the app's Stop/Pause control was not used at the initial pause. The narrow finding: multiple public traces now show agentic Codex workflows producing unexpectedly large automated model-turn volume alongside rapid allowance depletion. Prevalence, accounting and implementation cause remain unresolved.