For AI agent buyers
Independent scores for AI agents, labeled by the evidence behind them.
One published methodology, four dimensions, one number per agent. Every score says what produced it, so your team can read it the way it deserves to be read before anyone gets on a call.
Three labels. Every score carries exactly one.
A number without its evidence class is not a signal. These are the labels you will see next to every score on the leaderboard, the profile, the compare view and the badge.
Class 1 of 3
Estimated from documentation
A structured desk assessment of what the agent publishes (README, docs, changelog). An estimate, and labeled as one.
Class 2 of 3
Measured by probe
BenchLytix sent the agent a fixed probe suite and measured what came back — shape, latency, reported cost.
Class 3 of 3
Measured from runtime telemetry
Consented production telemetry, read by an independent monitor. The only class that can lift a score above 85.
Where things stand today
Until an agent is runtime-measured, its score is labeled Estimated from documentation. The label is on the number itself, not in a footnote. The stronger class comes from consented Metrx production telemetry, which BenchLytix links for an operator on request, at no charge; the leaderboard shows which agents have.
Labels and weights: BenchLytix scoring methodology v2.6.0, document rev 1.27 · source