Google adds agent telemetry and token quotas across BigQuery and Looker

Google Cloud said it has pushed a stack of AI agent observability features into general availability. The announcement is on the Google Cloud blog and is dated 13 November. Kate Grinevskala and Snigdha Jha, both Google Cloud product managers, wrote the post. The same post names features in preview, and an open-source eval tool called Prism.

The claim is one telemetry framework that covers first-party agents in BigQuery and Looker. It also covers custom third-party agents built with the Agent Development Kit or LangGraph. Google Cloud says the telemetry lands in BigQuery tables the customer owns.

What is in general availability

Managed conversation logging in BigQuery is now generally available, Google Cloud says. The Conversational Analytics API carries admin audit logs and optional data-read audit logs in Cloud Audit Logs. It supports VPC Service Controls, customer-managed keys and regional data residency, by Google Cloud's account.

Prebuilt dashboards also went to general availability. BigQuery Studio dashboards track sessions, top knowledge sources and latency. Looker System Activity dashboards track engagement and token usage for Conversational Analytics. Agent-to-agent workflows reached general availability as well, Google Cloud says.

FinOps controls for agent spend

The post names two controls for the cost of running agents. The first is big_query_max_billed_bytes, a per-query cap on scan bytes. It runs a dry-run estimate first.

The second is daily input and output token quotas, by project, user or agent. That one is in preview, Google Cloud says. The token counts recorded are input, output, thought, cached and total, on each agent call.

Prism, an eval tool

Google Cloud open-sourced Prism, an evaluation tool for Conversational Analytics agents. It is on GitHub at GoogleCloudPlatform/conversational-analytics-prism, is Apache-2.0, and was created on 3 November. The repo had 11 stars on 14 November.

Prism runs reusable test suites against agents grounded in BigQuery tables or Looker Explores. Assertions run as regex. Agents are discovered across regions. Google Cloud also names first-party agent evaluations in BigQuery, in preview, built on the Gemini Enterprise Agent Platform evaluation service.

A team already running agents on BigQuery or Looker can now draw a cost curve per user. It all sits inside one vendor's stack. A team on a different cloud can still read the shape of this release. Trace in, dashboard out. An open-source eval reads the same traces as production. The quota is in preview.

The one thing missing is a customer. The post carries no named customer and no quoted number from a live deployment. Every claim in it is Google Cloud's own. The eval tool and the token counts land where a buyer can measure them.

Source

Primary source: Enterprise-scale AI agent observability and governance with BigQuery and Looker on the Google Cloud blog, and the open-source GoogleCloudPlatform/conversational-analytics-prism repository.