Carmel Labs · AgentStatus · June 2026
The Two Failures Hiding in LLM-as-a-Judge
Calibration problems shrink with better technique. Competence problems do not. Why the dominant evaluation paradigm has a structural ceiling, and the two methods older than language models that get past it.
More research
Continue reading
July 2026
You Can't Measure What You Can't Reach
Reachability and reliability are different problems. Almost every tool solves at most one — valid validation needs both.
Read reportJuly 2026
"But My Agent Can Be Wrong and Still Pass"
Two kinds of wrong. We catch the one that gives itself away through instability and rule-violation, and we're honest about the residue no outside-in method catches without ground truth.
Read reportJuly 2026
The Residentiality Report
The datacenter blind spot: reachability from where users are — and the hard limit on what outside-in can claim.
Read reportMarch 2026
The State of AI Agent Reliability
We monitored 3,260 production AI agents across 48 countries. 89% with perfect uptime scored 0% on quality. The full data is inside.
Read reportApril 2026
The State of AI Agent Drift
88% of agents started giving worse answers at least once in 30 days. A look at how production AI agents drift - and the systemic March 29 event.
Read reportApril 2026
The Anti-Synthetic Monitoring Thesis
Datacenter vs residential reachability across 6,228 matched agents — access asymmetry, not correctness-by-origin.
Read reportResearch
9 Businesses You Can Build on Agent Behavioral Data
Insurance underwriting, credit scores, compliance certification, procurement intelligence - the commercial layer that sits on top of continuous agent monitoring.
Read report