Xiaomi's HarnesX: AI Agent Evolution Framework Unveiled

Xiaomi's research team has unveiled 'HarnesX,' a new framework designed to automatically improve AI agent harnesses. This approach can significantly boost AI performance without altering the core AI model itself. Here's the thing: A harness connects large language models to external environments, managing prompts, tools, and memory. Until now, these have been built and improved manually. HarnesX features a modular structure, allowing harnesses to evolve independently. Its core is an automatic optimization engine called AEGIS, which analyzes agent logs, plans improvements, generates code modifications, and blocks side effects. In tests, HarnesX improved performance in 14 out of 15 model-benchmark combinations, showing an average gain of 14.5 percent. For example, the open-source model Qwen3.5-9B saw a 44 percent improvement on an implementation planning benchmark. What's more, HarnesX supports a co-evolution approach, which simultaneously advances harness evolution and model training, leading to an additional average 4.7 percent performance gain. This co-evolution can only be applied to open-source models. The bottom line is this technology could make AI agents much more efficient and effective.