158M overnight tokens: delivery worked; the harness bill did not

Aashish Bhandari’s Sep 16 OpenAI Community write-up is an overnight factory receipt with the cost ledger still open. One evening: fix twelve design-review findings on a web service (transaction guard, retries, credential rotation, fragile handoff, thin diagnostics). Wall clock 13h 4min; ~5.5h active model time; 118 files / 16,369 lines changed; nobody typed code overnight. Morning verification was the point: a second agent from a different vendor rebuilt from scratch and reproduced 100/100 tests, 73/73 browser checks, and a byte-identical release package with no serious finding. Continuous unattended work can ship — if an independent judge exists.

The bill is the other half of the story. Models processed 158,137,319 tokens; 95.4% were re-reads of already-cached context, and only 0.42% was new text. Hold the image: roughly two hundred War and Peace copies read to write less than one. What already worked (and what failed under load) is harness engineering, not model heroics:

• Zero-token telemetry via Stop/SessionEnd hooks counted 55 components, 1,505 activations, and 1,299 tool calls for $0 model tokens — so the autopsy itself is not another full-context activation.
• Routing by work shape (companion note): parent on the strongest model; mechanical/verifiable work to cheapest; judgement to mid/top. List-price equivalent $99.80 vs $218.81 (54% cut) — and the independent verifier still passed.
• Idle-parent trap: a “no polling” rule held while the parent had its own work, then failed when idle — 89 waits (88 at 60s), 66 timeouts, 10.3M tokens / 6.5% of the run / 12.7% of list-price cost spent learning “not yet.” Claude Code’s harness can wake a parent on worker completion; Codex today cannot, so a declarative rule alone is not enough.
• Partial failure on worker reuse: four long-lived workers ate 49% of the run because follow-ups re-sent accumulated history every step.

Queued fixes are boring and correct: longer waits with a progress line, a zero-token watchdog script that wakes the parent only on completion/budget/stuck, one-worker-one-stage with fresh minimal context, and telemetry that counts waits instead of reconstructing them.

Why it matters for code review

Overnight harnessed momentum is real here — twelve findings closed, independent verification green — but peak-session mastery would have hidden the real tax: orchestration that re-sends the whole conversation to ask if a worker is done. Sustained factories compound only when continuity and verification are cheap relative to generation. FOMO that agents aren’t working while you sleep is answered by unattended runs plus independent rebuild gates and zero-token observability; without those, always-on agents just mint larger morning queues and burn tokens on “not yet.”