Google's Harness² Boosts AI Agent Performance 11.8 Points Without Touching the Model
Google Research open-sourced Harness², a framework that auto-tunes the prompts, tools, and memory around frozen models to boost agent performance on unseen tasks.
- Google Research released Harness², a framework for recursive agent harnessing with frozen model weights.
- Optimizes eight components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins.
- Reports up to 11.8 point gains on LAB, 11.5 on WorkBuddy-Bench, 4.5 on JobBench.
- Two recursion modes: parallel explores wider, sequential refines deeper via Propose–Probe–Compose steps.
- Supports GLM-5, Gemini 3.5 Flash, OpenCode and Codex executors across three professional benchmarks.
- Apache 2.0 licensed, Python 3.12, paper on arXiv, requires sandboxed network for honest evaluation.
Google’s Harness² optimizes the system around a frozen model
Google Research has released Harness², an Apache 2.0 framework that automatically revises the prompts, tools, memory, and orchestration logic surrounding a language model. Model weights remain fixed while the framework executes candidate revisions, compares their behavior, and carries supported lessons into later tasks.
An execution harness is the software layer that turns a language model into an agent. It supplies instructions, tools, stored context, delegation rules, and control flow. Harness² turns the manual process of tuning that layer into a repeatable search procedure, trading additional inference and execution compute for automated optimization.
Nested loops turn edits into evidence
Harness² separates optimization into task-level and harnessing-level recursion. The task-level loop uses a contrastive process called Propose–Probe–Compose:
- Propose: Generate candidate edits to the execution harness for the current task.
- Probe: Run those variants on the same task and record how their behavior differs.
- Compose: Build a refined harness from the observed contrasts between executions.
The refinement step requires no task-specific labels, reward signals, or ground-truth answers. Harnessing-level recursion then updates a persistent improvement harness with adopted configurations and evidence-supported lessons, giving subsequent tasks a stronger starting point.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
