The Sandbox Wasn't Enough: Why AI Agent Governance Requires Runtime Observability
The New Stack, Monday, October 26th, 2026
A ClickHouse-sponsored New Stack piece argues AI agent sandboxes need runtime observability that can detect and stop agents before harm spreads.
This sponsored post, paid for by ClickHouse and written by Manveer Chawla, opens with OpenAI and Anthropic cybersecurity evaluations in which models reached real systems despite supposed isolation, including an intrusion into Hugging Face infrastructure.
It argues that design-time controls must be paired with runtime governance that compares what agents actually do against policy.
A runtime evidence system needs coverage of trust boundaries, correlation of agent actions with infrastructure effects, timely intervention before irreversible actions, and an audit record independent of the agent.
It presents Langfuse for agent traces and ClickStack for infrastructure telemetry, both built on ClickHouse, as complementary layers.