The economics under most agent products are still hard to see. Teams ship an AI workflow without a firm cost-per-outcome understanding, reconcile the model bill against an average, and then quarter end arrives. A feature priced one way turns out far more costly to serve than expected. Why? A single run extends across a harness, one or more routers, and the tools it calls, and each one sees only its own slice. The harness knows the customer and features. The router knows the tokens. A tool can hold a sandbox open for four minutes or wrap a paid API, but those costs never reach the router. Every layer can be fully instrumented, and we’ll still have no single record of what the run actually cost, end to end. Today, we're proud to introduce AUDR (Agent Usage Detail Record), an open standard for that agent run cost record. The standard borrows from the telecom industry's Call Detail Record: meter usage now, price it later. Three rules make that possible: one run ID across every layer, one authoritative source per field, and strict merging with no silent overwrites. We're building AUDR alongside some of the best teams in the ecosystem, as spec reviewers, early adopters, or advocates, including @merge_api, @mastra, and @tinyhumansai, among others. For developers to get started easily, the spec plan includes adapters for key platforms and frameworks, including @nvidia NeMo Relay, Merge Gateway, Mastra, @vercel eve, @LangChain, @LiteLLM, and @OpenRouter Granular cost instrumentation is becoming foundational infrastructure for anyone building or monetizing agents. We think it should be open, neutral, and community owned. Read the spec, write an adapter, and help us improve it: openaudr.dev/