AWS has published a reference pattern for monitoring Codex use when developers run the coding agent locally with OpenAI models accessed through Amazon Bedrock. The design sends OpenTelemetry metrics to Amazon CloudWatch without placing a centralized proxy in the model request path.
In the pattern, each developer signs in with AWS IAM Identity Center and uses Codex on their workstation. A local OpenTelemetry collector listens only on the loopback interface, enriches metrics with organizational context, and publishes them to CloudWatch using AWS Signature Version 4.
The resulting dashboard can show active users, conversation turns, API requests, token usage, tool-call activity, models, teams, departments, and cost centers. AWS stresses that these are operational metrics, not a billing ledger, because actual costs can change with discounts, credits, and pricing adjustments.
The practical goal is governance. As coding agents move from trials to team-wide use, engineering leaders need to see adoption, reliability, and consumption patterns. The limitation is also clear: telemetry can guide decisions, but it does not by itself evaluate code quality or safety.