AHE-Beta ties agent harness evolution to user requirements
A research team proposes a requirement-centered loop for evolving agent harnesses, reporting higher task success and lower token use than existing methods on held-out tasks.
In plain words
An agent harness is the software scaffolding around a language model that lets it complete tasks. The paper describes AHE-Beta, a method that improves such harnesses by treating simulated user complaints as requirements, implementing them as general changes, and only keeping changes that fix a verified issue. On held-out tasks it reports higher success rates than several comparison systems.
What was announced
The authors argue that existing harness evolution methods rely on a single evolution agent guided by benchmark scores, which they say produces narrow patches and benchmark-specific local optima while ignoring implicit requirements. AHE-Beta instead runs a three-step loop: simulated users raise requirements as issues over multiple interaction turns, a team-based modification step implements them as generalizable changes rather than task-specific patches, and requirement-driven selection merges a change only after it resolves a verified requirement and passes review.
On held-out tasks, the paper reports pass@1 rising from 21.9% to 33.3% with 40.8% fewer tokens than its Base Harness, and higher pass@1 than Claude Code, Meta-Harness, and AHE. In a blind human A/B test, users preferred AHE-Beta in 31 of 48 valid pairs and rated it higher on all five dimensions. The authors also report that without further evolution the gains transfer to GLM-5.3 and DeepSeek-V4.1-Flash, with pass@1 improvements of 15.8% and 5.3% over Base Harness.
What it does not show
- The evidence is the paper's abstract only; the full methodology, experimental setup, and baselines could not be checked, and the reported results are the authors' own and not independently confirmed.
Sources
Source-support review completed 20 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 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.