News, releases, and notes from the Guren team.
October 26, 2026
We gave Claude Code the same five feature tickets on Guren and on Void. Pass rates matched and Guren cost about 1.4× as much, but when we removed one authorization check from both apps, Guren's audit, delivered by its Stop hook, decided the outcome: 18/18 against 6/18 on Haiku 5.5.
October 3, 2026
Guren is new to every model, so its conventions are taught through the guidance we ship. With it, agents take as many turns as on a hand-wired Hono stack at 1.26× the cost, and the top models pass every feature ticket, almost always through Guren's own features.
August 18, 2026
20 atomic tasks with hidden tests, three Claude models, with and without the harness guren agent:init installs: 360 runs. Same pass rate, 26–28% fewer turns, 12–25% lower cost.
August 14, 2026
We measured what the same feature costs an AI agent across six frameworks. For a framework the model has never seen, most of that cost turned out to be archaeology, not friction.
July 26, 2026
Why the Guren blog is built with Guren, what the framework actually does, and what I plan to write here.
