Enterprise AI programs are moving from prototypes into production faster than the tools used to inspect their decisions, data access and operating costs. Google Cloud, part of Alphabet Inc. (NASDAQ: GOOGL), has expanded agent-observability and evaluation capabilities across its Data Cloud, giving data and platform teams a more unified way to trace agent activity and monitor consumption.

The update covers first-party data agents in BigQuery and Looker as well as custom agents built with Agent Development Kit or LangGraph. Google Cloud says the framework captures telemetry, stores it in customer-owned BigQuery tables and lets organizations analyze that information through tools their data and operations teams already use. The practical objective is to turn an agent from a black-box application into a workload that IT can monitor, evaluate and govern.

For Google’s first-party data agents, Conversational Analytics can emit telemetry automatically. Queries routed to BigQuery data agents through agent-to-agent workflows can send Cloud Monitoring metrics, Cloud Trace spans and BigQuery conversation logs into the customer’s project. For custom agents, Google said a single-line SDK instrumentation option can stream structured telemetry into BigQuery.

The distinction between operational metrics and conversation content is central to the design. Google said aggregate measures such as request volume, latency and token use, along with sanitized first-party trace spans, exclude customer core content including prompts, model responses, internal thoughts and generated SQL. Organizations that need a deeper transcript record can opt into per-agent conversation logging, which writes prompts, responses and turn-level telemetry to a BigQuery table controlled by the customer.

Google also describes safeguards around that optional logging. The company said its logging checks skip a turn when a knowledge source uses row-level security, column-level policy tags or dynamic data masking, while chat execution continues if logging encounters an error. The source identifies support for Cloud Audit Logs, VPC Service Controls, customer-managed encryption keys and regional or multi-regional data residency for the Conversational Analytics API.

On the analysis side, BigQuery Studio offers dashboards for usage, active sessions, knowledge sources and latency. Looker System Activity dashboards track engagement and token use for Conversational Analytics, while user ratings and query review are in preview. Teams can also use an open-source reporting template or a Looker Block for more tailored monitoring, according to Google.

Cost governance is another focus. BigQuery conversation logs record input, output, thought, cached and total token counts alongside turn latency, the company said. Administrators can set per-query BigQuery byte-scan caps to block oversized work before it accrues on-demand scan charges. Daily input and output token quotas by project, user or agent are in preview for BigQuery and the Conversational Analytics API.

Google is pairing production telemetry with evaluation. Its Prism open-source tool can run reusable test suites for Conversational Analytics agents and stream completed evaluation runs into BigQuery alongside production logs. That combination matters because an agent that appears healthy at runtime can still regress after a prompt, schema or model update. Legacy monitoring often separated infrastructure metrics from AI-quality checks; Google’s approach places both in a warehouse and BI workflow, where enterprise teams can judge behavior, cost and reliability together.