ARA-Labs/Agent-Native-Research-Artifact

ARA‑Labs / Agent‑Native‑Research‑Artifact

What it is

• A lightweight, open‑source toolkit that lets AI‑driven “research agents” record every step of an experiment in a structured, verifiable format. It defines a file‑system layout (the ARA artifact) that ties together claims, code, configs, raw results and a searchable exploration graph.

Why it matters

• When agents run autonomously they leave behind a chaotic mess of overwritten scripts and terminal logs. ARA forces the agent to guard‑rail its output, crystallize the messy trajectory into reusable knowledge, and provide total observability through a concise visualisation. This makes the research reproducible, auditable and extendable without manually stitching together notebooks, PDFs and repos.

Core design principles

1. Guardrailing & Verification – every claim is wired to a concrete execution artifact and can be formally checked.
2. Crystallizing Insights – the system captures pivots, dead‑ends and rationales, turning them into first‑class nodes.
3. Total Observability – a generated HTML process map lets a human supervisor see the whole DAG without reading raw terminal output.

The toolkit – seven “Agent Skills”

Skill	What it does	How to invoke
research‑manager	Auto‑capture decisions, configs, ablations as you work	/research‑manager (or auto‑run)
compiler	Turn an existing paper/repo/notes into a full ARA artifact	/compiler <path>
rigor‑reviewer	Run formal checks that every claim is backed by execution	/rigor‑reviewer <dir>
research‑visualizer	Produce an interactive HTML map of the exploration DAG (or serve a live view)	`/research‑visualizer <ara‑dir> [--serve
research‑foresight	Answer “what should I try next?” with grounded, falsifiable suggestions	/research‑foresight <ara‑dir> "<question>"
context‑drop	Share a file/folder with another agent via a single URL (no repo needed)	/context‑drop <path>
research‑fuzzer	Treat the investigation like a fuzzer: predict next actions, flag loops, require self‑refutation before concluding	/research‑fuzzer

All skills are installed with a single one‑liner that reads wire-ara.md and registers the appropriate commands for the agent you are using (Claude Code, Cursor, Gemini CLI, Codex, etc.). Manual install is also possible via npx @ara-commons/ara-skills.

Artifact anatomy (the file layout)

example_artifact/
  PAPER.md                # short manifest (~200 tokens) – high‑level relevance
  logic/                  # *cognitive* layer – claims, experiment plans, design
    claims.md
    experiments.md
    solution/            # architecture, algorithm, constraints
  src/                    # *physical* layer – code, configs, environment
  trace/                  # *exploration* graph – DAG of attempts, dead ends
  evidence/               # raw tables, figures, extracted data

Cross‑layer bindings tie each claim to the exact experiment, code and evidence that support it; failed attempts are kept as first‑class nodes so agents never repeat the same dead‑end.

How to get started

1. From your project root, paste the one‑line command shown in the README. The agent fetches wire-ara.md, installs the seven skills and writes a routing map into its own context file (e.g., CLAUDE.md).
2. As you work, the research‑manager skill logs every decision automatically.
3. When you’re ready, run research‑visualizer --serve to inspect the full DAG or rigor‑reviewer to certify the artifact before publishing.

Compatibility

• Implements the Agent Skills open standard, so it works with any LLM‑powered coding assistant that supports that spec (Claude Code, Cursor, Gemini CLI, OpenAI Codex, GitHub Copilot, etc.).

Maturity & community

• The repository ships a CLI (ara‑skills), a demo repo (ARA‑Demo), and a full example artifact (examples/the-ara-of-ara).
• Accompanied by a peer‑reviewed arXiv paper (arXiv:2604.24658) that benchmarks the approach against a PDF + repo baseline.
• MIT‑licensed, with a CONTRIBUTING guide for adding new skills.

When to use it

• If you are running autonomous research agents (e.g., for ML model design, data‑pipeline discovery, or scientific simulation) and need a reproducible, auditable record of the process.
• When you want a single, portable representation that can serve as a conference submission, a shared research artifact, or a starting point for downstream agents.

---

Citation

@misc{liu2026humanwrittenpaperagentnativeresearch,
  title={The Last Human-Written Paper: Agent-Native Research Artifacts},
  author={Jiachen Liu and Jiaxin Pei and Jintao Huang and …},
  year={2026},
  eprint={2604.24658},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2604.24658}
}