AI Agent Cost Drift: Rolling Dashboards Miss 0.35%/Day Creep
Why do rolling-average dashboards miss a 0.35%/day AI agent cost increase? A math proof and 60-day test show the blind spot - and the fix.
An AI agent's input cost floor - system prompt, tool schemas, CLAUDE.md, MCP servers - can grow slowly every day without tripping fleet dashboards built on rolling baselines, because the baseline climbs right along with the drift. A new open-source tool, drift_anchor_gate.py, pins a frozen 'canary' snapshot at day zero and proves the blind spot mathematically: given window length w and alert threshold T, any uniform daily growth rate below g* = T^(2/w) - 1 is permanently invisible to rolling detectors.
Tested across six 60-day synthetic worlds, four common rolling detectors (rolling median, rolling mean, EWMA, and variants) never fired a single alarm on a context floor growing 0.35% per day - even though it grew 22.6% over two months. The frozen anchor caught the same drift by day 9, with zero false alarms on a flat, unchanged fleet.
The result cuts both ways: when workload per run doubles while the floor stays flat, the frozen anchor stays silent for all 60 days while the rolling detector correctly fires on day 30. The practical takeaway for engineering teams: recency-based baselines are structurally blind to slow, uniform cost creep, and pairing them with a fixed reference point is necessary, not optional.