Burning Codex Quota on a Multi-Model Agent Workflow, Then Hitting Thread Limits

craigbalding · x · 2026-09-16

The author ran a multi-agent, multi-model Codex workflow (ChatGPT Pro quota down to 14% after two resets) to transliterate Safeyolo's AI security proxy from Python to Rust. His model split: Astra xhigh as coordinator, Luna xhigh for coding, Astra low for dedicated verification, with explicit Luna xhigh overrides on two coding workers. The harness then hit agent-thread limits three times, forcing silent reuse of a Luna worker after source freeze and killing the independent reviewer — leaving expensive Astra xhigh doing 'narrow' checks. Lesson: don't silently reuse older higher-cost workers, but thread limits make cost-optimized agent splits fragile.

Related event: ChatGPT Pro User Burns Through 20x Quota After Two Resets(2 posts)→

Original post →

More from coding & agent

coding & agent channel →