Harness-Zero: Harness Distillation via Agent-as-Harness
A batch of arXiv papers treated the agent harness, memory and turn-level training state as first-class objects.
TL;DR
- One arXiv paper, Harness-Zero, framed agent harness distillation through an agent-as-harness approach.
- Companion papers proposed regularized recursive self-improvement of agent harnesses and a benchmark mapping the Pareto frontier of agent memory.
- Two more addressed reliability: Critical-State RL for diagnosing trainable states in multi-turn tool use, and an emergent-collusion study in long-horizon LLM agent interaction.
One arXiv paper, Harness-Zero, framed agent harness distillation through an agent-as-harness approach, and a companion proposed regularized recursive self-improvement of agent harnesses. [1] [2]
A third paper, DolphinBench, described mapping the Pareto frontier of agent memory, while Critical-State RL diagnosed trainable states for multi-turn tool use. [3] [4]
On the safety side, an arXiv paper studied emergent collusion in long-horizon LLM agent interaction, and on GitHub a project pitched a persistent cognitive memory layer for coding agents. [5] [6]
Why it matters
Treating harnesses, memory and training state as explicit design objects is what turns agent demos into systems that can be measured and improved.
Editor's note
All arXiv items are cited at the abstract level and were not peer-reviewed here; the GitHub memory layer is a self-described project page.