ARIS: Cross-Model Juries for Autonomous LLM Research

What it does ARIS provides Markdown-only skill files that orchestrate a research workflow inside Claude Code, Codex, Cursor, and other LLM agents. The agent follows a structured loop—plan, draft, adversarial review, iterate, and persist—so it can run literature reviews or experiments while you are away. There is also a standalone CLI, ARIS-Code, that supports multiple providers including DeepSeek, Qwen, and Doubao without requiring any specific editor.

The interesting bit The workflow splits roles into depth, breadth, and memory, using a cross-model jury so one model’s output is critiqued by another. The research wiki acts as persistent memory, meaning the loop can drive research but never acquit itself—accuracy is enforced by disagreement.

Key highlights

• Runs inside existing agents or standalone; no platform lock-in
• Extends beyond academia to due diligence, legal research, and market analysis via the ARIS-Anything spin-off
• Includes ARIS-Monitor, a macOS floating widget that flags which sessions are stalled waiting for user approval
• Recent releases hardened system-prompt handling to redact secrets and fixed SSE streaming edge cases after cross-model audits
• The same workflow generated ARIS-in-AI-Offer, a collection of 23 bilingual ML interview cheat sheets with derivations and PyTorch code

Caveats

• MCP tool dispatch is currently flagged as experimental in the release notes
• Documentation alternates heavily between Chinese and English, which can make scanning uneven
• The project shipped eleven CLI releases in one month to patch leaks and streaming bugs, so the surface area is still settling

Verdict Researchers and engineers who want structured, autonomous LLM workflows without vendor lock-in should look here. If you need a polished, stable SaaS with a fixed API, this is still a fast-moving open-source experiment.