AI

Harness^2: Recursive Agent Harnessing for an Open World

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 14/10/2026, 12:19:27 UTC.

Briefings

This article presents Harness^2, a recursive agent framework designed to automatically evolve execution harnesses for large language models. The study addresses two critical challenges in open-world settings: the lack of ground-truth verifiers for unseen tasks and the failure of task-specific RSI to generalize. Harness^2 employs two levels of recursion: task-level recursion uses a Propose-Probe-Compose step to guide edits without verifiers, while harnessing-level recursion extracts experience to update a continual improvement harness. Across three professional benchmarks and four model-harness configurations, the framework consistently improved GLM-5 and Gemini 3.5 Flash, achieving gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB.

Key points

• Harness^2 uses two connected levels of recursion: task-level recursion refines harnesses via Propose-Probe-Compose steps without ground-truth verifiers, and harnessing-level recursion extracts experience to update a continual improvement harness.
• Across three professional benchmarks and four model-harness configurations, the framework consistently improved GLM-5 and Gemini 3.5 Flash, yielding gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB.

Context

LLM agents' capabilities depend on their execution harnesses, and recursive self-improvement (RSI) offers a promising approach to automatically evolve these harnesses. However, deploying RSI in open-world settings faces two critical challenges: tasks are often unseen during harness design, leaving no ground-truth verifiers, and the process fails to generalize to novel scenarios. Harness^2 addresses these challenges through two connected levels of recursion.

Limitations

The study was conducted across three professional benchmarks and four model-harness configurations. Results are limited to the specific tasks and models tested, and the framework's performance gains may not generalize to all open-world scenarios or different model-harness combinations.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 19:15:07 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeihxzzlvf2jctgg5idhfkz7jj6j5a3udevbpdcmury6xha4h2ge42e