An overnight research colleague that cites real papers

What it does Luxas is a TypeScript harness for autonomous scientific research. You give it a topic in RESEARCH.md and it spends hours crawling literature across OpenAlex, arXiv, and paywalled venues, reading papers, designing experiments, and producing publication-grade figures. It then writes a LaTeX report, subjects itself to adversarial content and layout review, and emits a compiled PDF with genuine citations. The whole run is crash-recoverable and file-backed, with no vector store or embeddings in sight.

The interesting bit The architecture treats the filesystem as the single source of truth: brain state, sub-agent conversations, and memory all live in plain files, so a restart simply replays from log.jsonl. The experiment workflow is deliberately paranoid—tool_impl and tool_review agents are spawned blind to each other so the implementation cannot redefine semantics to pass its own tests, a failure mode the authors say they observed live.

Key highlights

• Multi-model by design: defaults to Claude and OpenAI o3, but a one-line env variable redirects the entire Anthropic family to DeepSeek-v4 or Kimi.
• Stateless harness built atop vendored pi-mono primitives, with deterministic finish-gates that no prompt can talk past.
• Literature crawler uses an anti-detect browser for paywalled venues.
• Experiment design spawns parallel implementation and test agents, using pytest as the only ground truth.
• Adversarial figure and typesetter loops block finalization until layout, palette, and caption integrity pass audit.

Caveats

• Not cheap: default Claude runs cost $20–80 per manuscript; the DeepSeek + Kimi dual profile drops this to $2–10 but sacrifices prompt caching.
• Requires local LaTeX and Python tooling, so it is not a browser-only toy despite the demo site.
• The README notes that some runs required restarts or pi_pushback.md iterations when the reviewer and brain genuinely disagreed.

Verdict Researchers who want an overnight literature survey with real citations and generated figures should look here; anyone hoping for a lightweight cloud API call should look elsewhere.