Google Research Releases Harness² Framework for Self-Improving AI Agent Systems

Recursion around the model

Google Research has released Harness², an open-source framework for improving the software systems that surround AI models. The project focuses on “recursive agent harnessing”: allowing an agent to propose, test and adopt changes to its own operating harness while leaving the model's trained weights unchanged.

A harness includes the components that turn a model into an agent capable of acting on a task. Harness² exposes eight such components for editing: the system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. Together they influence how context is assembled, what tools are available, how work is delegated and what happens during the model's action loop.

The framework connects two levels of iteration. At the task level, multiple candidate harness changes can be explored in parallel or refined sequentially against a task. At a broader harnessing level, the system retains adopted changes and evidence from completed tasks so that later work can benefit from earlier refinements. The stated goal is to handle open-world settings where new tasks were not known when the original harness was designed.

Google Research reports that one Harness² step improved results for GLM-5 and Gemini 3.5 Flash across three professional benchmarks and four model-and-harness configurations. Reported maximum gains were 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on the partial-pass measure used by LAB. The repository says larger task-level recursion budgets produced further gains. These are project-reported benchmark results and require matching configurations for meaningful comparison.

Open tooling with substantial operational requirements

The release supports OpenCode and Codex executors, along with JobBench, LAB and WorkBuddy-Bench. A local demonstration uses deterministic editors and real Python execution without model credentials. Full model-backed benchmark runs have a larger setup: Linux, Python 3.12 and uv are required, with Node.js, Google Cloud tooling and benchmark-specific dependencies needed for some configurations. WorkBuddy-Bench also requires Docker.

Harness² records task summaries, adopted harnesses, evidence and evaluation results under a runs directory, and repeated commands can resume a run. The documentation warns that judging and reporting must use the same benchmark, executor, task subset, recursion budget, mode and random seed as the refinement process. It also recommends dry runs before launching model-backed evaluations, which can incur API costs.

The repository flags an important evaluation risk: executors may have unrestricted network access while public benchmark rubrics can be available online. Users are advised to restrict networking so an agent cannot retrieve answers or grading criteria during a test.

Released under the Apache 2.0 license, Harness² is described as a research project rather than an officially supported Google product. Its main contribution is a concrete framework for testing whether systematic changes to agent infrastructure can deliver gains without retraining the model itself.