A new research paper from UC Berkeley, Northwestern University, and the University of Illinois Urbana-Champaign reveals that the "harness"—the system prompt, tools, and loop around an LLM—significantly impacts cost but not accuracy for AI agents. The study found that changing the harness can increase costs by up to 2.6 times for the same model and task, primarily due to the repeated re-reading of context in stateless LLM calls. Researchers analyzed traces from models like Claude Code, Codex, and GPT-6 Luna, observing that while accuracy remained similar across different harnesses, the more complex ones incurred substantially higher expenses, often due to verbose system prompts and tool schemas. AI
IMPACT Highlights that optimizing LLM agent harnesses can significantly reduce operational costs without sacrificing performance.
RANK_REASON Research paper detailing cost implications of LLM agent harnesses. [lever_c_demoted from research: ic=1 ai=1.0]
- Amazon Bedrock
- Claude Code
- GPT-6 Luna
- Northwestern University
- Opus 5.5
- Qwen3 8.27B
- Sonnet 5.5
- SWE-bench Verified
- System Optimizations for Cost-Efficient Agent Harnesses
- University of California, Berkeley
- University of Illinois Urbana-Champaign
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →