This AI system fixes its own broken tools mid-task, and smaller models benefit the most

Researchers at Xiaomi have built an AI framework that automatically rewrites the software layer connecting an AI model to its tools and environment. Ben Dickson reports for VentureBeat that the system, called HarnessX, treats this software layer, known as a harness, as something that can be modified and improved automatically rather than remaining a fixed, hand-coded setup.

Most AI agents today rely on static harnesses. Developers write them once and update them manually when something breaks or requirements change. HarnessX instead monitors how an AI agent performs during real tasks, identifies failures, and rewrites the relevant parts of the harness on its own.

What this means in practice

The researchers tested HarnessX across five task categories, including software engineering, web navigation, and multi-step reasoning. Key results include:

• An average performance gain of +14.5% across 15 model and benchmark combinations
• A +44% improvement for the smaller open-weight model Qwen3.5-9B on planning tasks
• An additional +4.7% boost when the underlying AI model was also retrained using data generated during harness evolution

One concrete example: during web navigation tests, an AI agent kept looping through search results without completing a purchase. HarnessX detected the pattern and inserted a mechanism that forced the agent to make a decision, resolving the loop and improving results.

The system uses a powerful frontier model, in this case Claude Opus, to act as the meta-agent that writes new harness code. This is a current limitation: open-weight models have not yet been tested in that role. Additionally, if the underlying task model is too limited for a given workflow, harness improvements alone cannot compensate.

For teams working with smaller, less expensive AI models on complex tasks, the findings suggest that improving the harness could deliver meaningful gains before investing in a more powerful and costly model. The researchers plan to release the code publicly.