A Cursor Agent Wiped a Database in 9 Seconds. Agent Analytics Would Have Seen It Coming.

TL;DR

• 1

  The 9-second database wipe is an analytics story, not just an agent horror story — Nate says normal dashboards would show an active user, long session, and lots of chat activity while completely missing the crucial signals: tool calls, permissions, retries, approvals, and where the human trust loop failed.

• 2

  The key unit of measurement is the agent run — instead of tracking page views or sessions, teams need a run-level view of what work was attempted, what tools were used, what broke, whether the task completed, and whether the user accepted the output.

• 3

  Chat logs and engineering traces are necessary but insufficient — chat transcripts are too qualitative to aggregate, and tracing tools capture latency, cost, and errors but not whether the outcome mattered to the user or improved the product.

• 4

  Completion and acceptance are different metrics, and the gap matters — a task can reach a finish state without earning user trust, which is why high completion plus low acceptance is a red flag that the agent is producing work users still have to redo.

• 5

  Salesforce is already shifting toward work-based metrics — Nate points to Salesforce’s February 2026 fiscal Q4 release, where it reported 2.44 billion agent work units growing 57% quarter over quarter, as evidence that major SaaS companies are moving beyond seats and sessions toward measuring AI-delivered work.

• 6

  The first three events every team should instrument are simple — track when runs start, when tasks complete, and when users interrupt or reshape the run, all tied to the same agent run ID so you can measure completion and correction rates by workflow.