SWE-Duel: coding agents duel by planting hidden bugs, 1,500-turn tournament ranks GPT-5.5 first
SingaporeManagementUniversity · hf · 2026-10-12
Singapore Management University introduces SWE-Duel, a self-scaling arena where coding agents evaluate each other via mirrored Red-Blue turns: the Red agent implements a feature and plants a subtle bug with sealed tests, and the Blue agent must spot and fix it blind. A 1,500-turn tournament across 15 repos in 5 languages ranks GPT-5.5 with Codex CLI first (98.3% defect repair) and Claude Opus 4.8 with mini-swe-agent as best task author (27.3% Red win rate). Newest-model-authored tasks beat a fixed solver 42.9% vs 4.0% for older versions, suggesting difficulty scales with model capability.
More from coding & agent
- Rive renders entire 3D city with vector graphics, ships CLI for Codex and Claude Code agents — soleio · 2026-10-12
- TechnoLama shows the AI-recreated PowerPoint lookalike from the ArtCraft suite — technollama · 2026-10-12
- Unpacking agent product forms: capability, collaboration and accountability are separate axes — sujingshen · 2026-10-12
- Dev vibe-codes a minimal Mac app that floats YouTube over everything — blakesamic · 2026-10-12
- Epic Ships Official Unreal Engine MCP, Letting AI Agents Drive UE Projects — Kyrannio · 2026-10-12
- AI-Generated 3D Movie Made Solely with Claude Opus 5.5 and BlenderMCP — sidahuj · 2026-10-12