AUDR: An Open Standard for Tracking AI Agent Runtime Costs

The Cost-Tracking Problem in the Age of AI Agents

When an AI agent completes a task, it may have called multiple large language models, triggered several external tools, and passed through different routing and billing systems. This raises a critical question: who initiated this run? And how much did each system it touched actually cost? Without a unified record-keeping standard, this information ends up scattered across different places, making it nearly impossible to consolidate or reconcile.

AUDR (Agent Usage Detail Record), launched by Chargebee, is a direct response to this pain point. It defines an open standard for capturing who initiated an agent run and the actual costs that run incurred across every system it touched. The project received 188 upvotes and 24 comments on Product Hunt, briefly ranking #2 in the Open Source, Developer Tools, and GitHub categories.

Inspired by the Telecom Industry's Call Detail Records

AUDR's design draws inspiration from CDR (Call Detail Record), a mechanism that has been battle-tested in the telecommunications industry for decades. In telecom networks, a single phone call may traverse multiple carrier networks, and CDRs provide a standardized billing and attribution record for each leg of that call — making cross-network settlement possible.

Similarly, a single agent run "travels through" multiple independent systems — execution harnesses, model routers, billing systems, and more. AUDR ports the CDR concept into the AI domain, defining a common JSON Schema that allows any framework, router, or billing system to emit and ingest these records.

The value of this analogy lies in what the telecom industry has already proven: as long as there's a unified record format and clear attribution rules, even the most complex cross-system billing can be tracked precisely. The AI agent ecosystem is still in its early stages, and a widely adopted open standard has a real opportunity to prevent the fragmentation that would result from every vendor building their own siloed, incompatible system.

---

CDRs have over forty years of practice in the telecom industry. Each CDR record typically includes the calling party, called party, start and end times, duration, network nodes traversed, and billing amounts. When an international call crosses multiple national carriers, each network segment generates its own CDR, with final settlement completed through standardized exchange protocols like TAP (Transferred Account Procedure). This mechanism made "usage-based billing with cross-network revenue sharing" a reality decades ago — without relying on bilateral private agreements between carriers. What AUDR aims to replicate is exactly this kind of settlement infrastructure built on "standard formats + clear attribution + mutual recognition," simply replacing "call minutes" with AI-native billing dimensions like model invocation counts, token consumption, and tool call fees.

Three Core Rules That Make the Standard Work

AUDR functions because of three core design rules:

Shared Run ID

Each run has a shared run ID generated by the execution harness. This unique identifier flows through every system the run touches and serves as the critical anchor for stitching together a complete cost picture from distributed records. Without a unified ID, cross-system data simply cannot be reliably correlated.

Clear Field Ownership

AUDR defines explicit field ownership, meaning it clearly specifies which system is responsible for populating each class of fields. This prevents the chaos that arises when multiple systems write to the same field and overwrite each other, making data provenance traceable and accountability well-defined.

Strict Merge Rules

At the data merging layer, AUDR enforces strict merge rules: conflicting records are rejected outright, and any corrections must appear as new records rather than in-place modifications to existing ones. This "append-only, never overwrite" design draws from best practices in auditable systems, guaranteeing record integrity and auditability — making it straightforward to trace the full history of every cost change after the fact.

What an Open Standard Means for the AI Ecosystem

As an open-source standard, AUDR's significance extends well beyond any single product. As enterprises increasingly deploy complex multi-agent, multi-model workflows, cost observability is becoming a hard requirement for both operations and finance teams. Who called what, and how much did it cost — these questions directly impact budget control, cost allocation, and customer billing.

A common schema adopted across vendors would allow tools from different providers to interoperate naturally, reduce integration overhead, and give third-party cost analysis, monitoring, and auditing tools a unified data foundation to build on. It's not hard to see the industry insight behind Chargebee — a company known for subscription billing — launching this standard. Metering and billing for AI usage may well be the next battleground in need of standardization.

Of course, the real value of any open standard depends on ecosystem adoption. Whether AUDR can become the "CDR of AI agents" hinges on how many execution frameworks, routers, and billing systems are willing to integrate with it. But from a design philosophy standpoint, it has identified the core of the problem: in an increasingly complex world of agents, transparent, auditable, and interoperable cost records are indispensable infrastructure.

---

Cost observability is an important component of the broader LLMOps (Large Language Model Operations) landscape. Today's mainstream observability tools — such as LangSmith, Helicone, and OpenLLMetry — typically focus on tracing call chains, latency, and error rates, while cross-system cost attribution remains a weak link. This is especially true when workflows span multiple model providers (OpenAI, Anthropic, locally deployed models) and multiple intermediary layers (like LiteLLM as a routing proxy), where no single tool can obtain a global view. AUDR aims to solve this at the data layer: as long as every system emits records in a unified schema, any downstream analytics tool can aggregate and allocate costs without needing to understand the details of upstream systems. This is quite similar to how OpenTelemetry is positioned in distributed tracing — not replacing each vendor's implementation, but providing a unified data contract.

Summary

AUDR offers an open answer to the challenge of tracking AI agent runtime costs through a concise yet rigorous set of rules. Drawing on the CDR model validated by the telecom industry over decades, it stands on three pillars — a shared run ID, clear field ownership, and strict merge rules — in an attempt to establish a widely adoptable standard before the agent ecosystem fully solidifies. For teams building or operating agent-based systems, this is an open-source project worth watching.