engineering
Loop Engineering in Production: Putting AI Agents on Call
We had an agent loop running in early 2026, and it still failed. Here is the scaffolding that made Sherlock, our investigation agent, useful in real incidents: 85% of the roughly 11,000 alerts a month suppressed as noise before a model ever runs, and investigations that once took 10+ minutes now finish in about 2 minutes on average, for less than a third of the cost.
