ARA-Labs/Agent-Native-Research-Artifact

What it solves

AI agents performing research often leave behind unobservable and unstructured data—overwritten code, scattered logs, and undocumented failures. ARA (Agent-Native Research Artifact) provides a structured, verifiable, and traceable format for AI scientists to document their work, ensuring that scientific claims are backed by ground-truth execution and that the research trajectory is preserved for human oversight.

How it works

ARA transforms research into a multi-layered artifact consisting of a root manifest, a cognitive layer for logic and claims, a physical layer for code and configurations, an exploration graph (DAG) that tracks the research journey including dead ends, and an evidence layer for raw proofs. It uses a system of cross-layer bindings to link claims to experiments and evidence.

To integrate ARA, users can "wire" it into their coding agents (such as Claude Code, Cursor, or Gemini CLI) using a single-line prompt. This installs a suite of six specialized agent skills:

• research-manager: Captures decisions, ablations, and configurations.
• compiler: Converts existing papers or repos into structured ARAs.
• rigor-reviewer: Verifies the epistemic rigor of an artifact.
• research-visualizer: Provides an interactive process map of the research trajectory.
• research-foresight: Allows users to ask grounded, falsifiable questions about the research.
• context-drop: Enables sharing files or folders with other agents via a URL.

Who it’s for

Human researchers and AI agents working together in "auto-research" environments where high-level oversight, reproducibility, and the ability to recover failure knowledge is critical.

Highlights

• Epistemic Anchoring: Every scientific claim is wired to ground-truth execution to prevent hallucinations.
• Preservation of Dead Ends: Failed approaches are treated as first-class nodes in the exploration graph to prevent agents from repeating mistakes.
• Total Observability: Translates complex agent behaviors into a minimalist interface for human course-correction.
• Agent-Native Standard: Compatible with the Agent Skills open standard and major AI coding agents.