Harness²

The core idea of Harness2 is to recursively rewrite the agent harness at test time without updating the model weights.

Harness²'s dual recursion brings computational overhead problems: each task requires multiple rounds of probing and composing, and the framework level also needs to maintain an experience library, placing higher demands on inference cost.

The task-level recursion uses a Propose-Probe-Compose loop to refine execution strategies.

The Propose-Probe-Compose loop does not require a ground-truth verifier, but instead relies on the observable results of the task itself for self-evaluation.

The framework-level recursion accumulates reusable experience across tasks, abstracting effective prompt structures, tool-calling patterns, and memory organization methods into transferable framework knowledge.

Google Research reported a performance improvement of up to +11.5 points on the WorkBuddy-Bench benchmark.

Experience distillation and context injection are the implementation foundation of the framework-level recursion; framework knowledge is stored in natural language form, does not change model parameters, and is instead dynamically concatenated into the prompts or context of subsequent tasks.

Continual learning based on context injection is limited by context window length, and when the experience library becomes too large, retrieval, compression, and filtering mechanisms are needed, otherwise noise or conflicts will be introduced.

Harness²'s unlabeled self-improvement depends on the quality and richness of environmental feedback, and may face challenges in scenarios where feedback is sparse or noisy.

Harness² received 108 votes on Product Hunt, placing third on the day's leaderboard.

Harness²'s Makers include researchers such as Han Yu.

Harness² uses a Dual recursion mechanism, running optimization loops at both the task level and the framework level.

Harness² (arXiv:2610.14224) pushes self-evolution to a second level: it can recursively edit the running harness during execution of the current task, and can also recurse at the level of "the harness that improves the harness".

Testing with two models from different families, Gemini 3.5 Flash and GLM-5, shows that the Harness2 method itself does not depend on specific model characteristics.

In the limitations section, the Harness² authors note that their signal for judging whether a change is effective does not guarantee correctness, nor does it establish that adopted changes will transfer to different tasks.

The Harness² paper reports improvements on three professional benchmarks.

The Harness2 code has been released on GitHub (google-research/harness2).

On October 9, 2026, Google Cloud AI Research published Harness², a framework that improves a model's surrounding systems without modifying the model itself.

A Google research team open-sourced Harness2, an Apache-2.0 licensed framework.