Operations | Monitoring | ITSM | DevOps | Cloud

Most teams judge their AI coding agent on two things: the monthly bill and a feeling. The bill tells you what you spent and the feeling tells you whether it seems to be helping, but neither one tells you what the agent actually did. As these tools move into the critical path of how software ships, that gap is starting to matter. I wanted to replace the feeling with something I could measure and understand what shapes of work affects this bill, so I decided to run an experiment on myself.

There was a time when “Opus” meant a classical composition and “Sonnet” was fourteen lines of Shakespeare you definitely did not read before the test. Now they’re model tiers, and every new release rewrites the economics of your engineering org whether you’re ready or not. Currently, your monthly total hides the crucial information you need to control and justify AI spend.

OpenTelemetry won the architectural standards battle. At scale, though, telemetry breaks more like plumbing than code. It breaks quietly, across a graph, with a blast radius you don’t understand until it’s expensive. With over 65% of organizations now running more than 10 collectors in production, hybrid deployments across Kubernetes and VMs are accelerating fast. Telemetry standardization is no longer a project milestone. It is a baseline expectation.

The race to better AI-assisted observability has been a race for bigger and better models. But intelligence was never the real bottleneck. Structure and context were.

Recently, someone talked Chipotle’s customer support agent into reversing a linked list – a task completely unrelated to burritos in any way. Screenshots circulated, people laughed, but underneath the joke sat a sharper question. If a production support agent will do that on a public channel, what else will it do that nobody is screenshotting? The bug is funny. The trust gap behind it is not.