← All cases

Agent orchestration

Idle AI quota turned into assets — 16 deliverables in one night

I measured that 42% of AI usage windows were sitting empty. A dispatcher started spending the leftover capacity on verifiable tasks, with numeric criteria and a reviewer that rewards refutation. First night: 17 tasks, 16 approved.

42%
of the 5h windows were empty (21 days measured)
16/17
tasks approved on the first night
−1,492
lines of legacy code removed by an agent
Client
In-house operation
Period
August/2026
Stack
Claude Code · Antigravity CLI · git worktree · cron · bash/python

The idea

A 5-hour window’s tokens are energy that can’t be stored. “Storing” means converting the leftover into durable assets: documentation, tests, audits, knowledge maps.

What was built

A scheduled dispatcher that reads the available quota on each account, only runs if there’s slack (fail-closed), executes one task at a time in an isolated worktree, with push and deploy made impossible by design, and logs cost per task and per model. Multiple accounts are rotated by fixed work type, never by whichever has the most quota left — because the provider blocks the whole set otherwise.

The rules born from real failure

  • Code only ships with a numeric acceptance criterion (one agent “met the letter” by copying 9,500 lines).
  • A cascading reviewer has to reward refutation — one that confirms everything invents a bug.
  • A cheap model that writes outside the isolation gets banned on the spot.

What this means for your business

Any company with an AI subscription is paying for capacity it doesn’t use at night. That capacity can become audits, documentation, and tests — backed by verification, not faith.