← back
What if the harness mattered more than the model? - Aditya Bhargava, Etsy
Takeaway
Invest in harness design and controlled tool access to improve agent capability independently of model upgrades.
Summary
- A cited HarnessBench comparison reports scores from 52.4% to 76.2% when changing harnesses under otherwise fixed evaluation conditions.
- Bhargava argues that better harnesses could make local open models more competitive and reduce reliance on proprietary frontier models.
- Agency is a new language for agent construction with automatic function-to-tool schemas and built-in execution logs.
- A median-function repair demo incrementally adds tools to a bare model call and introduces the need to constrain filesystem capabilities.
agent-harnessagencyopen-models
Original description
The models are getting so good now that all you need is a simple harness, some tools, and a loop... right? Aditya Bhargava argues that this direction promotes overreliance on large, proprietary models, and we should be focusing on building harnesses that improve the performance of models that can be run locally. Speakers: Aditya Bhargava (Etsy): Aditya Bhargava is a Staff Engineer and IC Initiative Lead of Agentic Commerce at Etsy. GitHub: http://github.com/egonSchiele