Google Research Releases Harness² for Recursive Agent Self-Improvement

Google Research has released Harness², an open-source framework for recursive agent harnessing designed to optimize the system surrounding an AI model rather than the model weights themselves. The framework operates on two levels of recursion: task-level recursion refines the execution harness for a specific task through a Propose–Probe–Compose step, while harnessing-level recursion updates a persistent improvement harness with lessons learned from previous tasks.

Core Mechanism

The framework treats the harness—comprising prompts, tools, memory, and orchestration—as the primary optimization target. It exposes eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. These components control context construction, tool interfaces, and behavior within the tool loop.

Task-level recursion works by proposing edits to the harness, probing their effects on the current task, and composing a refined harness based on behavioral contrasts between executions. This process occurs without ground-truth feedback. Harnessing-level recursion then retains the adopted harness and evidence-supported lessons to guide future refinements.

Performance and Benchmarks

Across three professional benchmarks and four model-harness configurations, a single Harness² step demonstrated significant improvements over base harnesses. Key results include:

• GLM-5 and Gemini 3.5 Flash showed gains of up to 11.5 points on WorkBuddy-Bench.
• Improvements of 4.5 points were recorded on JobBench.
• LAB saw an 11.8 percentage point increase in partial pass rate.

The framework supports both parallel recursion (exploring multiple candidate harnesses) and sequential recursion (refining deeper through successive steps). It is compatible with OpenCode and Codex executors.

Availability and Requirements

Harness² is available under the Apache 2.0 license. The release supports JobBench, LAB, and WorkBuddy-Bench. Users can install the framework via uv pip install -e '.[render]' and run benchmarks using commands like harness2 --bench jb --k 1 --mode parallel. The project is hosted on GitHub under google-research/harness2.