AI Agent Cost Spikes Expose a Missing Observability Layer

Usage-based billing made tool-using AI agent costs visible, but visibility isn't the same as explanation. When a team finds a run costing 40 times the median, provider meters show tokens and totals, application logs confirm success, and trace viewers show a tidy request-response — yet none explain why that particular run wandered while others finished cleanly.

The root issue is that agent observability is often built for the wrong unit of work. Traditional services handle discrete requests, but agents operate in loops: reasoning, calling tools, reading results, and continuing until they finish, hit a boundary, or escalate. Application-style tracing doesn't capture this looping structure, leaving control planes with policy opinions but no evidence to enforce them against.

The fix is a loop-aware observability substrate instrumented at unavoidable chokepoints — model gateways, tool proxies, policy engines — capturing per-turn token accounting, tool-call events, guardrail decisions, and identity context in a fleet-queryable store, with metadata captured by default and payload content only by exception. This turns cost from a mystery on an invoice into something engineers can actually attribute and govern.