Google Research Open-Sources Harness²: Agents Recursively Improve Their Own Framework Without Weight Updates

Google Research has open-sourced Harness², a plug-and-play framework for "recursive agent harness optimization" in open-world settings, released under Apache-2.0. The core idea: a harness is the surrounding system that turns a model into an agent—system prompt, tools, memory, skills, sub-agents, and so on. Harness² lets the agent automatically and iteratively improve this outer system while executing tasks, without any model training or weight updates, achieving gains of up to 11.8 points in a single round (also reported as 11.5 points).

Confirmed

• Released by Google Research; open-source repo at github.com/google-research/harness2 under Apache-2.0
• The optimization target is the full harness around the LLM: system prompt, tools, memory, skills, sub-agents, etc.
• Model weights stay frozen—no model training involved
• Performance gains reported across multiple posts, up to 11.8 points (m1) and 11.5 points (m4, m5)
• Project stars growing fast: m4 (10-18) reports 1.8k; m3 retweets say 2.7k

Why it matters

• Mainstream agent improvement relies on fine-tuning or RL to train model weights; Harness² shows a zero-training-cost alternative—recursive optimization of the surrounding system alone yields substantial gains, especially friendly to developers and teams with limited compute
• As a plug-and-play framework, it can be combined with existing agents, lowering the barrier to agent engineering and tapping into the recent community buzz around "context engineering / harness engineering"

Posts differ slightly on the maximum gain (11.8 vs. 11.5 points), possibly reflecting different benchmarks or rounds; refer to the paper and repo for exact figures.

2026-10-18 ~ 2026-10-19 · 6 related posts

• [source] Google open-sources Harness²: recursive agent harnessing lifts GLM-5 and Gemini 3.5 Flash by up to 11.8 points — dr_cintas · 2026-10-18
• [source] Google open-sources Harness2: agents that refine their own harness, up to 11.5-point gains — chemist_slime · 2026-10-19
• Google open-sources Harness², letting agents self-improve prompts, tools and memory with frozen weights — CurieuxExplorer · 2026-10-19

3 near-duplicate retellings: Saboo_Shubham_ · algo_diver · aigclink