We made claude​.ai 3x faster in two weeks. Here’s how we use Claude to measure, debug and improve performance. Prompts and methods included. claude.dev/blog/how-we-ma…↗

My job at @every is basically benchmarks-as-a-service now: every.to/also-true-for-…↗

An engineer on our team @thegp pointed out this is better than benchmarks - I agree - This kind of research into agents coordination, agent failure modes, and behavior in the face of false or contradictory information is much more insightful than the numerical scores we normally…↗

Benchmarks are dead (for us). Our RSI system makes benchmarks too easy. Give ours a benchmark and it builds its own solution–then beats SOTA. We just did it on 6 at once: math, coding, planning, long-context, tool use, web apps. No human tuning.↗

for those of you who are autoresearch pilled , or have been meaning to get into autoresearch but dont know how - I shipped evo today - a opensource Claude Code plugin that optimizes code through experiments you hand it a codebase. it finds a benchmark, runs the baseline, then fi…↗