Cost governors layer by time scale. Each layer catches a distinct failure.
time scale
layer
catches
per request (<1 s)
max_tokens
unbounded completion (single LLM call)
per session (min)
max_turns, max_budget_usd
long reasoning loop in one session
per tool (sess.)
per-tool cap (e.g., ≤5 WebFetch)
new tool triggers new loop (e.g. order-track)
velocity (10 min)
cut off above $50 / 10 min
runaway loop; fast-burn
per day
rolling daily cap + alert
slow leak; gradual doubling
per week/month
hard monthly cap + weekly alert
bad release; regression across sessions
orthogonal: cost reduction at constant quality
prompt caching · context windowing / compaction · tiered model routing · parallel cache for long-horizon runs
kill switch on breach
any cap fires → session aborts → alert → manual re-enable required → agent cannot modify its own cap
observed 2026: e-commerce agent $1,200 → $4,800 after tool addition; weekly-growth alert was the fix.