In October 2026 Google Research introduced Harness², a framework that lets agents improve the surrounding orchestration of prompts, skills, hooks, and other components (the "harness") around a frozen base model on previously unseen tasks.
Agent runs the base harness on each task to record a baseline trajectory.
Gemini 3.5 Flash +8.4 on WorkBuddy‑Bench and +7.5 on JobBench.
Skills and scripts drive ~50% of edits; dropping skills costs 5.3 pts.
The workflow consists of a base harness run, an Editor agent drafting two candidate edits, a Composer agent diffing and merging only edits that fix failures (subject to a configurable cap), and a Reflector agent logging verified lessons and storing winning components in a reusable library for future tasks.