The Google research team has recently made the agent framework Harness² available as open-source software under the Apache 2.0 license. This innovative framework significantly boosts the capabilities of AI Agents in a recursive manner, without the need to modify model weights. It achieves this by fine-tuning eight different types of external components, including system prompts, memory, and tool guidance.

Harness² operates on a dual-layer recursive mechanism, functioning at both the task level and the harnessing level. Remarkably, it does not necessitate tasks to possess clear, standard answers. Testing has demonstrated notable performance enhancements across a variety of professional task benchmarks. The framework provides both parallel and sequential recursion methods, allowing for an increased computational load during testing to attain higher success rates.

At present, Harness² serves primarily as a research project, offering support for OpenCode and Codex. Practical applications have shown that, upon integrating specific models, Harness² can markedly elevate performance in visual tasks. Industry experts highlight that employing Harness² necessitates adherence to the principle of 'model-harness-task matching,' emphasizing that there is no one-size-fits-all solution that is independent of the model.

Currently, related avenues such as AutoHarness are still in their nascent stages, with numerous practical issues—including evaluation, rollback, safety, and cost—requiring resolution. Nevertheless, several pertinent open-source projects have already been put into practice.