Ruiyao Xu - Harness²: Recursive Agent Harnessing for an Open World

​Can recursive self-improvement improve an agent’s harness, and how the harness itself gets improved? In an open world, agents encounter tasks their harnesses were never designed for, requiring adaptation at test time, often without ground-truth verifiers. This talk introduces Harness², a framework that addresses this challenge through two connected levels of recursion while keeping model weights fixed.

Task-level recursion improves the execution harness: the prompts, tools, memory, skills, and more that enable a model to act as an agent. Through contrastive Propose–Probe–Compose steps, it proposes edits, probes their effects on the current task, and composes a refined harness using behavioral evidence. This process can explore multiple candidates in parallel or build on successive refinements sequentially. Harnessing-level recursion improves the improvement harness: after each task, it retains adopted harnesses and evidence-supported lessons in an evolving library that guides future refinement.

The speaker will explain how these two levels connect improvement within a task with learning across tasks, and share experimental findings and examples, illustrating how agents can improve both their harnesses and the process of improving them.