My AI Agents Lie About Their Status, So I Built a Hidden Monitor

6 hours ago

The author built a hidden dashboard called 'Mission Control' to monitor 18 AI agents used daily for various tasks like Sales, Ad Analytics, and Support.
Initial attempts to have agents self-report their status via Google Sheets and HTML dashboards failed due to inconsistency and agent non-compliance.
Prompt engineering and persona-driven language adjustments did not improve agent reporting reliability.
The breakthrough came from leveraging Claude Cowork's built-in 'hooks'—infrastructure-level events that automatically track agent activity without requiring agent cooperation.
Three key hooks were used: UserPromptSubmit (for new tasks), PostToolUse (for in-progress tasks), and Stop (for completed tasks).
The final system intercepts agent activity data from shared directories and pushes updates to a centralized dashboard, eliminating reliance on agent self-reporting.
The solution highlights the importance of infrastructure-level automation over prompt-based instructions for reliable AI agent behavior.

Hasty Briefsbeta