Agent Telemetry Is the New Clickstream

This marks a shift in where AI infrastructure value accrues, from model calls alone to the exhaust those calls produce. Every agent run creates a stream of prompts, tool calls, traces, errors, eval scores, and cost records that teams need to store, filter, and inspect in near real time. That workload looks much more like web clickstream and log analytics than classic warehouse analytics, which is why ClickHouse has moved from product analytics into the default database layer for agent observability.

• ClickHouse was originally built for huge append only event streams like clicks and pageviews, so agent telemetry fits its core design. The common pattern is millions of small records written continuously, then queried fast by time, model, tool, user, or failure type while engineers debug bad runs and watch cost spikes.
  2 sacra 3 sacra
• In practice, this means ClickHouse is often the speed layer beside a warehouse, not a full replacement. At AstraZeneca, Databricks handled broader data processing, while ClickHouse handled agent retrieval and real time AI queries with sub 200ms latency on petabyte scale data, versus minutes on Databricks for similar dashboard style workloads.
  3 sacra 5 sacra
• The money follows observability economics. A prior ClickHouse operator described observability as the dominant workload, with materially better compression than OpenSearch, longer retention, and much lower cost, while another enterprise user said ClickHouse could be 10 to 20x cheaper than Datadog at scale. That makes agent telemetry a large new budget line, not just a feature inside existing tools.
  1 sacra 4 sacra 3 sacra

The next step is that agent databases turn into full agent operations stacks. ClickHouse is already moving from raw storage into managed observability, vector search, and agent facing products through acquisitions like Langfuse and HyperDX. As more software gains autonomous workflows, telemetry will become permanent system exhaust, and the winning platforms will be the ones that can query it cheaply and instantly.