Google Research has released Harness2, a plug-and-play framework designed to optimize the 'harness' surrounding AI agents rather than the models themselves. The tool addresses a critical gap in agent development: how to improve performance on tasks unseen during initial design without retraining expensive model weights. By treating prompts, tools, memory, and orchestration as an editable system, Harness2 enables recursive self-improvement in open-world scenarios.

The Mechanics of Recursive Harnessing

The core innovation lies in two levels of recursion that keep model weights fixed. Task-level recursion employs a contrastive 'Propose-Probe-Compose' step to refine the execution harness for specific tasks. It proposes edits, probes their effects on the same task, and composes a refined harness based on behavioral contrasts between executions, all without needing ground-truth feedback. Meanwhile, harnessing-level recursion updates a persistent improvement harness, storing adopted configurations and evidence-supported lessons to guide future refinements.

Benchmark Performance and Configuration

Early results across three professional benchmarksβ€”WorkBuddy-Bench, JobBench, and LABβ€”show significant gains. A single Harness2 step improved GLM-5 and Gemini 3.5 Flash performance by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB's partial pass rate. The framework exposes eight editable components, including system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins, allowing granular control over context construction and tool interaction.

Implementation and Infrastructure Requirements

Harness2 supports both parallel and sequential recursion modes, offering flexibility in how deeply agents refine their own execution strategies. Parallel recursion explores wider by probing multiple candidate harnesses, while sequential recursion refines deeper through successive steps. The release integrates with OpenCode and Codex executors, though users must configure their environments carefully, including Node.js 20+, Google Cloud CLI, and specific dependencies for benchmarks like Docker for WorkBuddy or Pandoc for LAB.

Key Takeaways

  • Harness2 optimizes the agent wrapper (prompts, tools, memory) without retraining model weights.
  • Uses a contrastive Propose-Probe-Compose method to refine harnesses without ground-truth feedback.
  • Achieved up to 11.5 point gains on WorkBuddy-Bench and 11.8 percentage points on LAB benchmarks.
  • Supports both parallel (wider exploration) and sequential (deeper refinement) recursion modes.
  • Licensed under Apache 2.0, but explicitly noted as not an officially supported Google product.

The Bottom Line

This is a pragmatic shift from brute-force model scaling to surgical harness engineering, letting developers squeeze more juice out of existing models like Gemini 3.5 Flash without touching the weights.