Google and Northwestern's Harness² self-improves agent harnesses without answer keys
rohanpaul_ai · x · 2026-11-15
A Northwestern/Google paper shows agent harnesses (prompts, tools, skills, memory around a model) can self-improve on unseen tasks without answer keys. Harness² runs the agent, has an editor model propose harness edits, then reruns tasks to keep only changes that improve behavior; lessons accumulate into a library guiding future edits. A single step lifts GLM-5 by up to 11.5 points on WorkBuddy-Bench (past Gemini 3.5 Flash on its stock harness), +4.5 on JobBench, +11.8pp on LAB — while blind editing without test runs drops coding scores from 73.4 to 63.2. Open-sourced at github.com/google-research/harness2.
More from coding & agent
- arXiv: Harness^2 — recursive agent harnessing for open-world tasks — rohanpaul_ai · 2026-11-15
- Dev builds a PowerShell AI harness for Microsoft's new Intelligent Terminal — dfinke · 2026-11-15
- snitch: a harness that watches agents in parallel and reverts bad paths — generativist · 2026-11-15
- Dev: unattended agent runs burn 99% of tokens on useless reproducibility gates — generativist · 2026-11-15
- workspace-cli: open-source Rust CLI for AI agents to drive Google Workspace — majidmanzarpour · 2026-11-15
- Multimodal voice-driven harness control plane source shared alongside a simple ADK tutorial — gerardsans · 2026-11-15