We gave our AI agent fleet a credit limit, and it hit it the same day

Description and use

A fleet of ten autonomous AI agents was capped at $100 over a rolling 5‑hour window. The system treats the shared event log as a transaction log, replays it into double‑entry ledgers, measures and prices labour, alerts on budget thresholds, and automatically throttles new task dispatch until spend falls below the cap. This provides real‑time cost control without manual intervention.

Problem

Uncontrolled AI agent fleet spending exceeded budget, risking runaway costs and loss of control.

Outcome

The fleet’s spending was capped at $100 over a rolling 5‑hour window, automatically pausing new tasks when the cap was exceeded and resuming as spend aged out. The system provided real‑time audit, alerts, and control without manual intervention.

Workflow

• 1. Replay board events into three hledger journals (money, promises, labour).
• 2. Verify ledger parity with hledger check and independent replay.
• 3. Query outstanding obligations via mesh-promises.
• 4. Measure spend with mesh-labor over a rolling 5‑hour window.
• 5. Price spend using token counts and a rate table.
• 6. Alert on budget thresholds with mesh-labor-alert (edge‑triggered).
• 7. Throttle new task dispatch with mesh-pace when spend exceeds cap.
• 8. Allow operator overrides via Telegram and direct command channels.
• 9. Auto‑resume dispatch as spend ages out of the rolling window.
• 10. Update budget cap in environment file and re‑source when needed.

Evidence

The board gets replayed into three separate double‑entry hledger journals, each tracking a different commodity: money, promises, labour.

Source

mesh-labor --budget … spent: $43.23 / cap $200 / remain $156.77 …

Source

mesh-pace holding, new dispatch stopped … operator raises the cap to $200 → spend $152.38 < $200 → dispatch resumes

Source