Per-Agent Cost Tracking for Multi-Agent AI on AWS

This article explores the critical challenge of observability in multi-agent AI systems, where traditional monitoring often fails to capture silent financial waste. Even when an agentic workflow returns a successful result, it may be consuming excessive tokens or performing redundant reasoning cycles that inflate costs. The author demonstrates how to implement per-agent cost tracking on AWS using the Traccia SDK and OpenTelemetry. By instrumenting spans with specific attributes like token counts, agent identity, and dollar costs, developers can gain granular visibility into which agents, tools, or workflows are driving expenses. The guide provides a practical, read-only approach to building an 'AWS Account Investigator' crew, allowing developers to identify and mitigate silent overspending without modifying production infrastructure. This methodology moves beyond basic uptime monitoring, enabling teams to enforce cost governance and detect performance regressions in complex, autonomous AI environments.