Glite ARF runs 12 parallel coding agents with verifier-driven research

Researchers from Glite have introduced Glite ARF (arXiv 2606.27416), an open-source Python framework that tackles a core problem in automated empirical research: naive delegation to LLM coding agents leads to broken, irreproducible artefacts at scale. The framework defines a three-role architecture: a human researcher selects hypotheses, coding agents (Claude Code, Codex CLI) implement structured tasks, and deterministic Python verifier scripts enforce task isolation, immutability, materialized overviews, and a corrections overlay. This "verifier-driven research" approach ensures rules live in code that fails loudly, not in prose agents merely follow.

To demonstrate its effectiveness, the team applied Glite ARF to the BEA 2026 vocabulary-difficulty shared task, achieving first place in the closed track and second in the open track across Spanish, German, and Mandarin. The campaign reduced the official baseline RMSE by 29.9% (closed) and 35.9% (open) using 273 tracked tasks (146 experiment runs) across 129 feature sets, all orchestrated from a single laptop with up to twelve parallel agents and some model training on rented A100s. Total third-party cost was under $500, including approximately $450 in LLM API spend. Crucially, the framework's per-fold provenance detection caught and stripped four target-leaking feature sets, correcting an implausible 0.609 RMSE to 0.802. Across three campaigns, the structural machinery added only about 1% wall-clock time overhead. The framework and a public demo project are available under Apache-2.0.