harness2

Harness²: Recursive Agent Harnessing (Google Research / arXiv:2610.14224). Dynamically adapts and self-improves the agent execution harness on-the-fly using dual-level recursion (Propose-Probe-Compose

SKILL.md

GitHub

Harness² — Recursive Agent Harnessing for Open-World Tasks

Based on Google Research (arXiv:2610.14224), Harness² enables recursive self-improvement (RSI) for LLM agents without retraining or updating underlying model weights. It operates by iteratively diagnosing, adapting, and evolving the execution harness (tools, prompts, memory, validators, and middleware).

Dual-Level Recursive Architecture

Level 1: Task-Level Recursion (Propose – Probe – Compose)

When facing an open-world task where ground-truth automated verifiers may not exist:

  1. Propose: Generate candidate repairs or modifications to the current harness and plan.
  2. Probe: Run lightweight diagnostic probes and contrastive execution traces to assess viability and side effects.
  3. Compose: Synthesize successful probe elements into an upgraded execution harness tailored to solve the immediate task.

Level 2: Harnessing-Level Recursion

  • Extracts generalized insights across task-level runs.
  • Updates the agent's long-term harness library without accumulating noise or catastrophic forgetting.
  • Evaluates whether new harness patterns (guardrails, tool definitions, verification heuristics) improve performance on benchmark suites.

How to Apply Harness²

  1. Contrastive Trace Analysis: When an agent hits an execution error, contrast the failing run against previous successful patterns.
  2. Harness Mutation: Adjust prompt constraints, tool parameters, or verification gates dynamically.
  3. Probe Validation: Validate the mutation on a sandboxed sub-problem before committing to the full task.