Evolving Agent Harnesses for Domain Specialization
arXiv — cs.AI ·
What to know
The paper proposes evolving the agent harness (system prompt, tool schemas, skill library, planning scaffolding) while keeping the policy frozen, using GEPA, a reflective evolutionary optimizer. Experiments with Codev on SWE-bench Verified, OfficeBench, and Terminal-Bench show harness evolution recovers most gains of weight updates at a fraction of the compute, and combining both is best. The authors argue harness adapters avoid failure modes of SFT-based specialization like shortened reasoning chains.
Source and evidence
arXiv — cs.AIOriginal language · und ·Original author: BoHao Chen · Source: arXiv and the named paper authors. Descriptive metadata: CC0 1.0.Source licence
Why it matters by industry
- Artificial intelligence
Relevant to AI models and agents, as it introduces a compute-efficient alternative to weight updates for domain specialization, applicable to coding agents and potentially other embodied or multimodal systems.
Reviewed summary. The complete article remains at its original source.