Writings

Latest thinking
from Sikaru.

Engineering, research, and ideas behind agents that improve from real work.

2 articles

FrontierHarness pass rate against cost per task on Kimi K3: Sikaru at 86.7% and $1.84, MiniMax Code at 76.7%, Codex at 66.7%, with the published-harness frontier from Exo Harness to Codex.

EngineeringOctober 30, 2026

Self-improving managed agents, steered by your team

An amber optimization path climbs from a local minimum toward a higher peak on a green wireframe metric landscape.

AnnouncementsOctober 13, 2026

Introducing Sikaru: self-improving agents, with the lab built in

SikaruSelf-improving agents, with the lab built in.