Luxas¶
external · status: active · focus: end-to-end · discipline: general · started: 2026
Project page: https://github.com/Muuuun/luxas
Licence: MIT
Source: projects/landscape/luxas.yml
Positioning¶
An unattended question-to-manuscript harness: a Claude Opus "brain" agent coordinates 13 specialised sub-agent kinds (search, reader, worker, experiment, tool_impl, tool_review, experiment_reviewer, math, illustrator, illustrator_write, typesetter, reviewer, fixer) through literature discovery over OpenAlex/arXiv/CrossRef, experiment design, parallel implementation and testing, figure generation, LaTeX writing and a three-layer adversarial review, ending in a compiled PDF. The entire pipeline runs from a one-line topic in RESEARCH.md with no human gate — it sits at the far-autonomy end of the RISE end-to-end layer, next to aris, agon and autoresearchclaw.
Distinctive contribution¶
Two structural anti-self-deception mechanisms rather than prompt instructions: implementation and testing are written by separate blind agents (tool_impl works from specifications only, tool_review writes test suites without seeing the implementation), and the finalize loop cannot be bypassed by prompting — a figure-internal audit (palette, spines, typography, clipping) and a document-level layout audit via PDF rasterisation gate completion, with a \resultref mechanism tying every claimed number to a computed result. It is also the most actively developed system in this lane (382 commits, 981 stars, pushed 2026-09-06) and publishes nine complete example runs as inspectable PDFs.
Evaluation scores¶
| Dimension | Score (0–3) | Note |
|---|---|---|
| Lifecycle coverage | 3 | Twelve declared stages from topic decomposition through literature survey, experiment design/implementation, symbolic derivation, figure generation and LaTeX drafting to adversarial review and a compiled PDF — end-to-end including both drafting and a validation stage. |
| Autonomy level | 3 | One-line topic in RESEARCH.md plus luxas run produces a compiled manuscript over multiple hours with no approval points; the README's own framing is 'while you sleep'. |
| Architectural transparency | 3 | MIT with the full TypeScript harness in src/ (382 commits), the 14-role agent roster documented in the README, plus skills/, schemas/, patches/ and four benchmark task specs; no separate prompts/ directory (agent prompts live inline in src) and no scoring harness for the benchmark tasks. |
| Inputs supported | 1 | A single input form — one free-text topic — combined with literature access via OpenAlex, arXiv, CrossRef and citation chains; no dataset input, no private-corpus connector, and experimental data is generated by the agent rather than supplied. |
| Outputs / reproducibility | 2 | Persists LaTeX source, compiled PDF, figures, generated Python, and a full log.jsonl replay trail with \resultref number provenance; but runs are stochastic and no manifest or seed makes a given manuscript reproducible from its inputs. |
| Internal evaluation | 1 | Nine gallery runs and four benchmark task directories (bench-01-u233-gamow, bench-02-rb-blockade-leakage, bench-03-ftqc-ccz-conflict, bench-04-prior-art-positioning) to inspect, but no paper, no scoring harness, no reported metrics and no third-party evaluation. |
| Openness | 2 | MIT-licensed with a documented install path (Node 22+, LaTeX, poppler, tmux), but a run costs $20–80 on the default full-Claude profile ($2–10 on --profile dual), so the demonstrated examples are not reproducible on free infrastructure. |
| Maturity / traction | 2 | ~1,170 stars (up from 981 in early September), 18 forks, 382 commits, pushed 2026-09-06 — the most active project in this lane — and now also offered as a hosted browser demo at luxas.im alongside the self-hosted CLI; still single-maintainer, version 0.1.0 in CITATION.cff, and no tagged GitHub releases. |
| Cross-family policy | 1 | Cross-family is configurable (--profile dual runs DeepSeek-v4-pro text with GLM-5.3-flash vision, OpenAI o3 backs the math agent, Kimi via env var) but the default profile keeps executor and adversarial reviewer inside the Claude family. |
| Runtime assurance | 3 | Blind separation of tool_impl from tool_review so the agent cannot self-validate circularly, a reviewer deliberately held outside the brain, three-layer adversarial review (content, figure internals, PDF layout), deterministic finish-gates that prompting cannot bypass, and \resultref number provenance — a full runtime audit stack with gating. |
| Cross-platform portability | 2 | Four-plus model providers behind env-var redirects (Anthropic default, DeepSeek, GLM vision, OpenAI o3, Kimi) but a single execution environment — its own Node/tmux harness, explicitly not a general agent framework and not usable inside another IDE. |
Scored on 2026-10-01. See the evaluation rubric.
Tags¶
Pipeline stages: rq-formulation hypothesis-generation literature-discovery literature-synthesis research-design formal-modeling code-generation data-analysis paper-drafting revision-editing referee-simulation dissemination
Architectural features: multi-agent tool-use iterative-loop persistent-memory
Inputs: research-topic
Outputs: paper-draft compiled-pdf figures analysis-code run-logs
Data sources: self-generated-experiments
Knowledge sources: openalex arxiv crossref web-search
Limitations¶
- Uses an anti-detect browser to reach paywalled academic venues — a publisher terms-of-service and licence-compliance exposure that should be assessed before institutional use; the README does not document what the evasion does or offer a compliant-only mode.
- All demonstrated work is physics: the nine gallery runs and all four benchmark tasks cover quantum error correction, Rydberg atoms, cold atoms and neutral-atom imaging, so performance on economics or social-science research designs is entirely unevidenced.
- No paper, no benchmark scores, no external validation — quality evidence is nine self-selected example runs.
- Costs $20–80 per run on the default profile and runs unattended for hours; the README warns that project directories should be treated as executed code with no credential isolation guaranteed.
- Single maintainer, version 0.1.0, no tagged releases — bus-factor and API-stability risk.