I burned through a weekly Codex allowance this week without ever asking Codex to write a single line of code. It was wired into exactly one role in my agent setup: the advisor. The second model that reads over the first one's shoulder and occasionally says "careful".

Not even a heavyweight, either. GPT-5.6 Luna — the variant OpenAI sells for "fast, high-volume work", the one with the highest message limits of the family. That's the one I ran out of.

Five days of logs: the advisor made 4,561 requests. The model doing the actual work made 2,795. The reviewer outspent the worker by 63%.

The reason is boring and structural. The advisor reviews every step, not every task — 131 things I asked for turned into 2,774 agent steps, and every one of them got reviewed. Its context is append-only, so each review re-sends everything it has already seen. 86,000 tokens in, twenty tokens of thought out, 4,561 times over. Call it 390 million tokens in five days to have someone mutter over my shoulder.

Best part: of 1,498 attempts to advise me, 383 actually made it into my session. The rest were dropped as duplicates or noise — after the request was already paid for.

So I switched it off. Running the agent bare now, no second model watching, to find out how much of that advice I was actually getting value from. Maybe the harness stands on its own feet. Maybe I find out on day three what the reviewer would have flagged on step two.

Only one way to learn what a safety net was doing, and that's to take it away.