UC Berkeley Introduces RHI: Optimizing Agent Harnesses to Cut Inference Costs by 60%
ceciletamura · x · 2026-08-06
Researchers from UC Berkeley and other institutions have proposed Recursive Harness Self-Improvement (RHI), a novel method for optimizing agent harnesses under model-harness co-evolution.
- Core Mechanism: RHI represents the harness as a prompt-level specification of the agent loop and iteratively refines it using pairwise feedback over its own revision history.
- Results: Across 30 synthetic ML research tasks spanning quantitative finance, robotics, and pharmacy, a few RHI iterations substantially raised the performance ceiling of low-reasoning-effort agents.
- Efficiency: This approach not only exceeded the corresponding maximum-reasoning-effort settings but also reduced inference costs by up to 60%.
- Key Insight: The performance gains arise primarily from improved inter-agent information flow and task-specific context management, rather than simply relying on longer reasoning traces.
More from coding & agent
- Hermes Desktop Integrates Hindsight Memory Provider — max_paperclips · 2026-08-06
- Most AI Agents Are Just Orchestrated Workflows — aftahi_ai · 2026-08-06
- Developer Slams GPT 5.6 for Chaotic Coding: Generates 30+ Files But Fails Basic Writing — obinopaul · 2026-08-06
- LLMs Outpace Humans: Automated Workflows Are Inevitable — teortaxesTex · 2026-08-06
- Long-Running Agent Memory: Avoid Auto-Compaction — repligate · 2026-08-06
- Orbital: An Open-Source Project-Level AI Agent — Aggravating-Risk1991 · 2026-08-06