← News

Agents & MCPSources: arXiv · 20 Nov 2026

arXiv paper proposes evolving agent harnesses instead of model weights for domain specialization

A new arXiv paper argues that adapting an agent's system prompt, tools and scaffolding can recover most of the gains of weight-based fine-tuning at a fraction of the compute.

In plain words

AI agents are usually customized for a domain by retraining the model, which is expensive. The paper instead evolves the surrounding setup — the instructions, tool definitions and planning structure — while leaving the model unchanged. The authors report this cheaper approach recovers most of the retraining gains, and combining both works best.

What was announced

The paper, posted to arXiv, frames domain specialization of LLM agents as more than a weight-update problem. The authors study evolving the agent harness — the system prompt, tool schemas, skill library and planning scaffolding — while keeping the policy frozen, casting this as an RL-style outer loop instantiated with GEPA, a reflective evolutionary optimizer. They say the resulting harness adapters need no GPU training, can be loaded per vertical domain, and compose with weight-based RL.

In experiments with Codev, a Claude Code-style coding agent, on SWE-bench Verified, OfficeBench and Terminal-Bench, the authors report a consistent pattern: weight updates gave the strongest single-path gains, harness evolution recovered most of those gains at a small fraction of the compute, and combining both was best overall. They also describe an SFT failure mode — shortened reasoning chains and increased truncation — which they argue harness adapters avoid by construction, positioning the approach as a non-destructive complement to weight updates.

What it does not show

  • The evidence is the paper's abstract only; the full experimental setup, baselines and results have not been reviewed, and the reported gains are the authors' own claims, not independently confirmed.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.