Routing tasks to specialized AI-agent harnesses reportedly beats Meta-Harness in four test settings

A post describing a new paper says two AI tuning branches each kept the practice problems it handled better and notes on what worked. One learned to verify math answers; the other learned to build full derivations. A router then selected a harness for each task, reportedly beating Meta-Harness in all four test settings. A harness is code around an LLM that controls its tools, retrieval and self-checks.