Harness²

Google Research has open-sourced Harness2, a framework that gives AI Agents' "scaffolding" the ability to self-improve.

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

The core finding of Harness² is that a smaller, weaker model, with the aid of a self-evolving harness, fully surpasses a stronger frontier model running on a default pre-built harness in performance.

Harness²'s two-level recursive mechanism and its weak-model-upset results have already attracted attention from the research community.

Harness² is still a research project at present.

Harness² adopts a two-level recursive mechanism: Level 1 evolves the harness in real time for the current task, and Level 2 evolves the way harnesses are built for future tasks.

The recursion at Level 2 is essentially meta-learning, i.e., learning the meta-knowledge of how to build a suitable harness, equivalent to an outer loop.

On complex tasks such as legal red-line annotation and budget reconciliation, single-step evolution raised accuracy by 35 to 50 percentage points.

Testing a candidate harness online requires 4 to 20 full Agent runs per task, so Harness² is not suitable for real-time use on one-off prompts and is better suited to accumulating a domain library over the first 10 to 20 tasks.

A core finding of Harness² is that a smaller, weaker model running in a self-evolving harness can completely beat a stronger frontier model running in a default harness.

Harness² proposes that an Agent should edit at four layers instead of adding rules to the system prompt: on-demand skills, custom tools, sub-Agents, and editing its own control loop.

Researchers had AI blind-write the harness (making changes without testing first), and performance dropped 10%.

Harness² uses a three-step Propose-Probe-Compose loop to evolve the harness without an answer key.

Squared refers to two layers of recursion: the first layer evolves a harness for the current task; the second layer evolves how the system builds harnesses for future tasks and stores reusable tools, skills and experience in an experience library.

In coding, web, office and legal scenarios, the runtime-evolving harness beat the static harness in every domain, including unmodified Claude Code and Codex.

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.