Dabo

Dabo is a self-evolving agent harness that uses the hardware itself as the feedback signal. You type a GPU task in chat, link your own harness from GitHub, and pick the devices: a Jetson Orin Nano (CUDA), a Mac mini M2 Pro (MLX/Metal) and Android phones with Adreno GPUs (OpenCL). A coding agent writes and tunes the kernel on the device. A locked judge scores it on hidden inputs with its own clock and a physics check. A coach (Claude Opus 5.5) then proposes one change to the harness itself: its rules, context policy, guardrails, tools or limits. A gate trials the change on the device and keeps it only if the judge's score goes up. Each device grows its own harness lineage, and a shared layer carries what helps everywhere. The model never changes, so every gain comes from the harness.