Xiaomi's HarnessX framework rewrites its own AI scaffolding mid-task

What happened

Xiaomi researchers introduced HarnessX — a framework in which the agent harness (the set of tools, prompts, and control logic) is itself a composable object the model can rewrite mid-task.

Context and impact

Until now, agent harnesses have been static: a developer designs them up front and the model just uses them. HarnessX flips the dynamic: the model can add a tool, rewrite its system prompt, or change policy the moment it hits a limit. The most surprising result: counter to intuition, small models (1-7B) benefit significantly more than frontier models, because they leverage the harness as 'external memory'. It fits into the post-Claude Skills/Plugins and OpenAI Codex Skills trend — but goes further by making the skill itself runtime-modifiable.

Details

• Harness components: tools, prompts, control flow — all named and editable
• Self-rewriting: model emits special tokens that modify its own scaffolding
• Biggest gains: small models (Qwen 3 0.6B – 7B); frontier models ~5-10% lift
• Use cases: long-horizon coding, multi-step research, Pokémon-style gameplay
• Trend: post-Skills frameworks (Anthropic Skills, OpenAI Codex Skills)