Muuuun/luxas

What it solves

Luxas is an autonomous research agent designed to handle the entire scientific research pipeline from start to finish. It solves the problem of manual literature review, experiment design, and report writing by automating the process of crawling academic papers, running computational experiments, and producing a publication-grade LaTeX PDF report with real citations and figures.

How it works

Luxas operates as a stateless harness that coordinates a multi-agent system. It uses a "brain" agent to decompose a research topic and sequence tasks, which then spawns specialized sub-agents for specific roles:

• Literature Survey: Agents crawl OpenAlex, arXiv, and CrossRef, reading and extracting key information from papers.
• Experimentation: A three-phase workflow (Design $\rightarrow$ Implementation + Review $\rightarrow$ Integration) ensures code quality. Implementation and testing agents work blindly to each other to prevent self-circular failures.
• Verification: An adversarial review process involves a PI (Principal Investigator) agent, a figure auditor, and a typesetter that rasterizes the PDF to check for layout issues.
• State Management: All state is file-backed (using log.jsonl and Markdown files) rather than relying on in-memory embeddings, making the system crash-recoverable.

Who it’s for

It is intended for researchers or students who need a literature survey or a small-scale computational study and want a fully automated, reproducible, and auditable deliverable in the form of a compiled manuscript.

Highlights

• End-to-End Automation: From a single prompt to a compiled LaTeX PDF.
• Adversarial Review: Three layers of self-review covering content, figure internals, and PDF layout.
• Blind Implementation: Separates code writing from test writing to ensure rigorous validation.
• Multi-Model Support: Native support for Claude, DeepSeek, GLM, and OpenAI o3.
• Crash-Recoverable: Stateless design allows the system to resume from file-backed logs.