Google's Harness² recursively self-improves agent harnesses, up to +11.8 points
burny_tech · x · 2026-10-20
A Google Research paper, Harness^2: Recursive Agent Harnessing for an Open World, improves agent execution harnesses (prompts, tools, memory, skills, orchestration) recursively while keeping model weights frozen.
- Task-level recursion: a contrastive Propose–Probe–Compose step — propose harness edits, probe their effects on the same task, compose a refined harness from behavioral contrasts — guides edits reliably without ground-truth verifiers
- Harnessing-level recursion: distills experience from adopted edits into a persistent improvement harness that guides future refinement
Across three professional benchmarks and four model-harness configs, a single step consistently beats base harnesses: GLM-5 and Gemini 3.5 Flash gain up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, and 11.8 percentage points on LAB, with further gains as recursion scales.
Related event: Google Open-Sources Harness², a Recursively Self-Improving Agent Framework(3 posts)→
More from coding & agent
- Hugging Face ships gr.workflow and calls on the open-source community to build with it — erhannah · 2026-10-20
- TraceVista: An Open-Source Tool for Visualizing and Understanding Agent Trajectories — _AustinCalvert_ · 2026-10-20
- Why GRPO collapses on multi-turn agent training, and how to stabilize advantage estimation — burkov · 2026-10-20
- Clairvoyance turns a C++ strategy game into a pixel-perfect web app for rapid idea testing — draginol · 2026-10-20
- Claude Code finds hidden camera in Airbnb smoke detector via Wi-Fi scan in 2 minutes — 4KTV · 2026-10-20
- SF meetup 'Are We Running Too Many Agents?' hits capacity with exe, Amp, DatologyAI speakers — josh_wills · 2026-10-20