An Agent-Native Research Artifact (ARA) is a research knowledge package whose primary reader is an AI agent, not a human. Instead of a single paper, it stores the research as structured, machine-readable layers: the current best understanding (
logic/), the full journey that produced it — dead ends included (trace/), and grounded proof (evidence/). Each demo here also ships a human-facingPAPER.mdand a self-contained interactivetrajectory.html.
Click the image to open the interactive trajectory viewer.
An AI agent (Claude Code / Opus) played the ARC-AGI-3 game ls20 (Locksmith) with no
tutorial and cleared all 7 levels. The ARA in arc-agi3/ls20/ was built
live, during play — and the agent then used it as a world model to crack the levels it
could not solve cold (L4, L6, L7). An ablation on the unseen L7 showed a recipes-only agent
never operated a single lock control, while the world-model agent characterized the whole
lock: the ARA earns its keep as a world model, not a trajectory cache.
Click the image to open the interactive trajectory viewer.
An ARA compiled from an autonomous agent's optimizer search on the fixed-architecture
modded-nanogpt track_3_optimization benchmark. Starting from the Muon baseline
(3500 steps to 3.28 validation loss), the agent produced statistically-validated records at
3205 → 3037 → 2949 steps (v1/v2/v3), plus a documented terminal negative result from a
hard-isolated novelty-constrained wave.
Browse everything online on the ARA Hub page for this repo — the tables above link straight to the live viewers.
Everything is also static and self-contained — you can clone the repo and open the same HTML
files directly in a browser, or read the ARAs as plain Markdown: start with a demo's
PAPER.md, then dig into logic/ (what is known), trace/ (how it was learned, dead ends
and all), and evidence/ (the proof).
| Layer | What it holds |
|---|---|
PAPER.md |
Human-facing summary of the artifact |
logic/ |
Current best understanding: claims, concepts, problem, solution recipes |
trace/ |
Append-only research journey: exploration tree, session records, raw notes |
evidence/ |
Grounded proof: verbatim numbers, figures, tables |
trajectory.html |
Self-contained interactive visualization of the whole process |
- Agent-Native-Research-Artifact — the main repository: ARA specification and the agent skills that produce, query, and visualize ARAs
- The Last Human-Written Paper: Agent-Native Research Artifacts — the paper introducing ARAs
- ARA Hub — browse published ARAs online
- aracommons.com — ARA Commons