OpenAI Codex Allegedly Spawns 826 Agents, Bills $78K

This is a serious accountability failure if substantiated: autonomous agent systems need hard concurrency, spend, and audit controls before they can be trusted with unattended work.

• –The report’s local evidence suggests model and reasoning escalation, recursive task expansion, and unusually high token amplification.
• –A separate Codex issue documents extreme token growth and repeated compaction involving parallel subagents and mixed 0.144.x versions, indicating a potentially broader failure mode.
• –OpenAI’s own product positioning emphasizes parallel multi-agent workflows, making transparent limits and kill switches essential.
• –Local counters are not authoritative billing data; only OpenAI can reconcile task IDs, server-side usage, invoices, and credit reloads.
• –Developers using agentic coding tools should enforce spending caps, concurrency limits, approval gates, and durable external logging.