Google Cloud now provides a single telemetry pipeline that records activity from native Conversational Analytics agents and custom agents built with the Agent Development Kit or LangGraph, writes the data to BigQuery tables you control, and surfaces it in dashboards across the Data Cloud. The change matters because it turns opaque agent execution into a traceable, auditable, and cost‑visible process that can be integrated with existing monitoring and security tooling.
What changed: unified agent observability
The new framework works the same way for every agent type. It captures Cloud Monitoring metrics, Cloud Trace spans, and conversation logs, then lands them in BigQuery tables owned by the agent’s project. First‑party agents emit telemetry automatically; custom agents require a single‑line SDK call to stream the same structured data.
Why it matters to AI, platform, DevOps, and security teams
- Traceability: Multi‑step reasoning can be followed end‑to‑end, helping AI engineers debug prompt chains and verify answer quality.
- Cost awareness: Token consumption and query spend appear in the same tables used for other operational metrics, allowing SREs to set alerts and budget controls.
- Governance: Admin audit logs and optional data‑read audit logs are emitted to
Cloud Audit Logs, giving security engineers a source of truth for compliance checks. - Integration: Because the data lives in BigQuery, existing Looker dashboards or custom analytics pipelines can be reused without additional data movement.
Architectural and operational considerations
Implementing the framework requires provisioning a BigQuery dataset that will serve as the “governed home” for all agent telemetry. The dataset inherits the project’s IAM policies, and you can apply VPC Service Controls to restrict network egress. For encryption, the tables support customer‑managed encryption keys (CMEK), and the service respects regional or multi‑regional data residency settings.
From an operational standpoint, the telemetry flow adds a lightweight write path to each agent invocation. Teams should monitor the write latency and storage growth, especially if high‑volume agents are in production. Existing monitoring dashboards can be extended to include the new metric streams, and alerting policies can be built on cost or error thresholds directly from the BigQuery view.
Security and governance implications
The managed conversation logging feature is GA and includes always‑on admin audit logs. Optional data‑read audit logs can be enabled to capture who accessed the stored conversation records. Because the logs are stored in a customer‑owned BigQuery dataset, data residency and encryption are under the customer’s control, reducing reliance on implicit cloud‑side protections.
Security teams should verify that the BigQuery dataset’s access controls align with the principle of least privilege, and that any downstream analytics tools inherit those controls. Enabling VPC Service Controls adds a network‑level boundary that can prevent exfiltration from the telemetry dataset.
Related CloudNinjas coverage: Google Cloud.
What This Means For Practitioners
- Enable the managed conversation logging feature in your Gemini Enterprise project to start receiving telemetry automatically for first‑party agents.
- For custom agents, add the single‑line SDK call provided by the ADK or LangGraph documentation to stream metrics and traces into your BigQuery tables.
- Create a dedicated BigQuery dataset, apply appropriate IAM roles, and consider
CMEKandVPC Service Controlsto meet your security posture. - Extend existing monitoring dashboards to include the new metric tables, and set alerts on token usage or error rates.
- Review audit‑log configurations to ensure admin and data‑read events are captured for compliance reporting.