Every production agent raises the same two questions, and most teams still answer both from guesswork. What did the agent actually do on that request, and what did the request cost?
On November 14, 2026, Google Cloud announced expanded agent observability and evaluation across Google Data Cloud. First-party agents in BigQuery and Looker now emit telemetry with no setup, and custom agents built on Agent Development Kit (ADK) or LangGraph stream the same structured data into a warehouse you own through a single line of SDK instrumentation. The same announcement moved the Conversational Analytics API and BigQuery conversation logging to general availability.
Primary source: New AI agent observability and governance across BigQuery and Looker (opens in a new tab).
What shipped
The release has three layers: capture, store, and analyze.
| Capability | Status | What it gives you |
|---|---|---|
| Conversation logging in BigQuery | GA | Prompts, responses and turn-level telemetry written to an agent_events table you own |
| Conversational Analytics API | GA | Always-on admin audit logs, optional data-read audit logs, VPC Service Controls, CMEK and data residency |
| BigQuery Studio agent monitoring | GA | Prebuilt dashboards for usage, active sessions, top knowledge sources and latency |
| Looker System Activity for Conversational Analytics | GA | Engagement plus estimated input and output tokens across BI workloads |
| BigQuery Agent Analytics plugin for ADK and LangGraph | Available | Full tool-invocation execution trees, time-series telemetry and multimodal payloads |
| Daily input and output token quotas | Preview | Caps per project, user or agent, set in the console and the gcloud CLI |
| Prism | Open source | Reusable evaluation suites with regex assertions that stream results into day-partitioned tables |
Agent-to-agent workflows that route queries into BigQuery data agents through Gemini Enterprise are now GA as well, and they emit Cloud Monitoring metrics and Cloud Trace spans into the agent's own project with no extra wiring.
Every logged turn carries the fields a spend review needs:
turn = {
"input_tokens": ...,
"output_tokens": ...,
"thought_tokens": ...,
"cached_tokens": ...,
"total_tokens": ...,
"latency_ms": ...
}
Two controls sit on top of that data. A per-query byte scan cap (big_query_max_billed_bytes) uses dry-run estimates to block oversized queries before they run, and daily input and output token quotas can be enforced per project, per user or per agent.
Why it matters: telemetry that lands in your own tables
Most agent stacks today log to a vendor console. The turn happens somewhere you rent, the trace lives outside the tools you query with SQL, and the cost arrives as one number on an invoice at the end of the month.
Google's change moves the record of the work into infrastructure the customer already governs.
one agent turn
|
v
[ Cloud Monitoring ] aggregate metrics, no prompt content
[ Cloud Trace ] step-by-step spans, sanitized
|
v
[ agent_events ] prompts, responses, token counts
| (your table, your IAM, your CMEK)
|
+--> BigQuery Studio dashboards
+--> Looker System Activity
+--> AI.GENERATE summaries, AI.EMBED drift clustering
The privacy design is worth reading closely. Operational metrics and sanitized trace spans deliberately exclude prompts, model responses, internal reasoning and generated SQL. Conversation logging is opt-in per agent, and a fail-closed check runs before each turn: when a knowledge source uses row-level security, column-level policy tags or dynamic data masking, the turn is skipped rather than logged. A logging error is fail-open, so chat sessions keep running.
That split is the part platform teams have been asking for. Metrics and traces by default, transcripts only where policy allows, and both governed by the dataset's IAM, CMEK and VPC Service Controls.
How CommandLyne handles this work
CommandLyne applies the same principle one layer up, across every model and every tool a team runs.
Governance and multi-model control (opens in a new tab) routes each task by complexity, role, latency and cost, keeps Bring-Your-Own-Key across Gemini, Claude and OpenAI, and reports token-level cost so a spend review has a real number in front of it. Security and enterprise memory (opens in a new tab) adds scoped tool access, permission-aware personal and team memory, human sign-off checkpoints and a full audit trail, with VM-level isolation per user and customer-managed encryption through KMS or CMEK.
Observability and governance are the same story told at two altitudes. Google gives a data agent a trace and a token count inside BigQuery. CommandLyne gives the workflow that calls five systems a route, a cost and an approval record.
Ad hoc rollout compared with a governed one
| Question | Ad hoc agent rollout | Governed agent rollout |
|---|---|---|
| What did the agent do? | Screenshots and guesswork | Trace spans and opt-in conversation logs in a table you own |
| What did it cost? | One invoice line at month end | Input, output, thought and cached tokens per turn |
| Who can read prompts? | Unclear, often the vendor | Opt-in per agent, fail-closed on masked data |
| What stops a runaway query? | Nothing until the bill lands | Per-query byte caps and daily token quotas |
| Which model ran this step? | Whoever wired the last script | Routing rules per task, role, latency and cost |
| Who approved the write action? | Nobody remembers | A named approver in the audit trail |
Where to start
- Turn on Cloud Monitoring and Cloud Trace in the agent's project, then opt in to conversation logging for one high-value agent.
- Add the BigQuery Agent Analytics plugin to your ADK or LangGraph agents so custom work lands in the same
agent_eventstable as first-party traffic. - Set a per-query byte cap and a daily token quota before the next traffic spike, not after it.
- Run the Prism suite against every prompt and schema change, so an accuracy regression shows up in a benchmark rather than in a customer complaint.
Try CommandLyne to route every agent step to the right model and see its cost per turn, or book a demo with our team.
Try CommandLyne (opens in a new tab) | Book a demo (opens in a new tab)

