6 Claude Code agents burned tokens on idle loops until one exit fix cut costs 6×

Icy_Comfort_6220 · reddit · 2026-07-21

A small zero-human publishing setup running 6 Claude Code agents found that the biggest cost driver was not model price, but orchestration mistakes and idle loops.

What went wrong

The orchestrator used a frontier model for simple triage.
Heartbeat runs kept waking up, finding nothing to do, and failing to record a clean exit.
That caused re-wakes and repeated empty runs, which quietly burned tokens.

What changed

Right-sized the models: routing/triage agents moved from Sonnet 4.6 to Haiku 4.5, while the Writer and Researcher kept the stronger model.
Added an early-exit gate: if no work is assigned, the run exits immediately before doing deeper checks.
Used prompt caching for the large static system prompt.
Added per-agent budgets with soft alerts.

Cost impact

Orchestrator heartbeat: about $0.40 → $0.06 when idle, or $0.11 when it actually works.
A typical idle heartbeat now costs $0.06 and exits in 8 seconds.
The writer remains the biggest cost line item, but that is acceptable because it is the actual product.

The core lesson: the real problem was frequency × giant prompts × idle loops, not token price alone. A cheaper model would have hidden the bug instead of fixing it.

Original post →

6 Claude Code agents burned tokens on idle loops until one exit fix cut costs 6×

More from coding & agent