Google Cloud and Cornelis strengthen monitoring and networking of AI agents
Google Cloud adds supervision, evaluation, and cost control tools for its AI agents. Meanwhile, Cornelis launches CN6000, an 800G network family designed to accelerate AI and HPC clusters.
On November 13, 2026, Google Cloud introduced new monitoring, evaluation, and budget control tools for its AI agents, with safeguards on sensitive data and dashboards in BigQuery and Looker. At the same time, Cornelis unveiled CN6000 on November 12, an 800G network family intended for AI clusters and high-performance computing. Two different announcements, but the same common thread: making AI systems more observable, faster, and a bit less surprising for the teams operating them.
When the AI agent becomes a service to closely monitor
At Google Cloud, the novelty lies in a clear separation between operational telemetry and conversation content. Aggregated metrics go through Cloud Monitoring, while Cloud Trace tracks execution steps without exposing prompts, responses, the model's internal thoughts, or generated SQL. In other words, the administrator sees the engine, not necessarily everything said inside the cabin.
Google-managed Conversational Analytics agents automatically emit this telemetry. Custom agents created with the Agent Development Kit, or ADK, and LangGraph can also send structured data with one-line instrumentation. All data lands in BigQuery tables owned by the client, with their own permissions, client-managed encryption keys, and VPC Service Controls.
This choice matters for companies wanting to industrialize their agents without losing governance control. Google also notes that conversation logging remains opt-in, agent by agent, and skips some conversations when the knowledge source uses row-level security, column policy tags, or dynamic data masking. If a logging error occurs, the chat session continues. The system thus avoids turning a monitoring tool into a secret vacuum, which would have been a bit too convenient to be reassuring.
From dashboards to billing, sometimes just one query away
The same BigQuery layer is also used to manage costs. Each conversation turn records volumes of input, output, thought, and cached tokens, as well as total latency. BigQuery Studio provides ready-to-use dashboards on usage, active sessions, most queried knowledge sources, and latency. Looker offers System Activity dashboards for Conversational Analytics, showing engagement and token usage.
Google Cloud also adds consumption limits. BigQuery queries can be blocked before execution thanks to caps on bytes scanned, calculated in dry-run. Daily quotas for input and output tokens are in preview for BigQuery and the Conversational Analytics API, applicable per project, user, or agent. The message to teams is clear: monitoring is no longer only after the fact; you can also shut off the tap before overconsumption.
Evaluate before deploying, then keep results handy
Google Cloud also highlights Prism, an open-source tool that tests Conversational Analytics agents connected to BigQuery tables or Looker explorations. It allows building reusable test suites, writing assertions by regular expression, and verifying response quality before a prompt, model, or schema change goes into production. Results can be written into daily partitioned BigQuery tables alongside production logs.
A native evaluation of BigQuery agents is also available in preview via an API and a guided Colab notebook. In this architecture, teams can compare the behavior of a test and production agent in the same data warehouse, with the same access controls. It’s cleaner and especially easier to audit when an agent starts answering a bad question with a lot of confidence.
CN6000, or how to talk faster between machines
Cornelis tackles another bottleneck: the network between servers. CN6000 is a family of 800G network adapters and switches designed for AI and HPC workloads, with Omni-Path and Ethernet versions. The company announces up to 1.6 billion messages per second on its Omni-Path SuperNIC, a bidirectional and aggregated figure that alone does not indicate if a model will truly be faster in production.
The manufacturer also says it measured 0.95 microseconds latency in an internal 8-byte MPI ping-pong test, compared to 1.2 microseconds for NVIDIA InfiniBand XDR at 800G, about 21% less in this specific context. The important point is the nuance: bandwidth, message throughput, and latency do not measure the same thing. A network can excel at small exchanges without magically speeding up a poorly designed distributed job.
CN6000 also integrates RoCEv2 over Ethernet, credit-based flow control, adaptive routing, two 400G data planes, as well as security features like secure boot, equipment attestation, and post-quantum signatures. Cornelis indicates the hardware is available for evaluation and initial deployments, with partner systems expected in Q1 2027. For buyers, the real question remains the actual workload: scientific computing, disaggregated inference, KV cache transfer, congestion, network software — everything plays out in the details.
Key points
- November 13, 2026: Google Cloud extends monitoring and quotas.
- November 12, 2026: Cornelis launches CN6000 at 800G.
- BigQuery becomes the common foundation for client logs.
- Prism compares agents before production changes.
- Cornelis announces 1.6 billion messages per second.
- CN6000 partners expected in Q1 2027.
By the numbers
- 0.95 microseconds — CN6000 latency in an internal 8-byte MPI test.
- 1.2 microseconds — NVIDIA InfiniBand XDR latency in the same test.
- 800G — capacity announced for CN6000 as of November 12, 2026.
- 1.6 billion messages per second — bidirectional aggregated capacity announced by Cornelis.
- Q1 2027 — availability of partner systems announced by Cornelis.