Benchmarks are the new stars

GitHub stars broke as a credibility signal because clicks cost nothing to fake. Frontier labs replaced them with benchmark scores. But benchmarks are already being gamed. The market is climbing from stars to benchmarks to verified outputs — each rung costs more to fabricate than the last.

April 10, 2026 · 6 min · mercurialsolo

Anatomy of a Voice Agent

Humans hand off a conversational turn in about 200 ms; production voice agents take 1.4 to 1.7 seconds. What I learned about that gap building menta, a clinic front-desk agent, and Solo, a meeting avatar.

June 2, 2026 · 24 min · mercurialsolo

HydraBench: Agent Infrastructure Resilience

23 scenarios, 5 claims, 345 measured runs on OpenHydra (460 planned across 4 frameworks). HydraBench tests what most agent benchmarks ignore: does your infrastructure survive crashes, contain secrets, deliver handoffs, enforce permissions, and control cost?

February 23, 2026 · 4 min · mercurialsolo