Codex Astra Exposes AI’s Token Waste
An AI engineer’s Codex postmortem found repeated context, unchanged-code reviews, and scope creep consuming two weeks of allowance in days. He now matches tasks to direct execution, bounded delegation, or capped review-and-repair loops.
More capable models make orchestration discipline essential: the real metric is useful work per token, not raw model activity.
- –Audit token usage by completed task, model, implementation, coordination, and review.
- –Reuse evidence when repository state is unchanged instead of replaying full context.
- –Route simple documentation fixes directly and reserve delegation for genuinely independent work.
- –Use acceptance criteria and bounded review loops to prevent endless polishing.
- –Cross-model reviews should focus on repair diffs and unresolved findings, not unchanged code.
DISCOVERED
1h ago
2026-09-14
PUBLISHED
1h ago
2026-09-14
RELEVANCE
AUTHOR
logicalicy