The authors formalized harness evolution as a multi-objective optimization problem, using a set of weight vectors to define scalar subproblems across the Pareto frontier, optimized jointly within a shared population.
Multi-objective evolution yields stronger, more transferable agent harnesses
Harmony Harness decomposes objective space and factorizes harness components to outperform single-objective monolithic approaches by over 10% on secure-coding benchmarks.
Big Tech
Weichen Yu · Lesly Miculicich · Bhavana Dalvi Mishra · Jiefeng Chen · Shuo Li · Matt Fredrikson · +2 more
Google Cloud AI Research · Google DeepMind · Carnegie Mellon University
Research Digest··2 min read
Thread:Agent Harness Optimization
The authors introduce Harmony Harness, a framework that treats LLM agent harness evolution as a multi-objective optimization problem using decomposition-based evolutionary search.
Why this paper
From Google Cloud AI Research and 2 others · Part of Agent Harness Optimization, now 135 papers
In one line
Harmony Harness jointly optimizes functionality and safety in LLM agent harnesses, outperforming single-objective baselines by over 10%.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§