RUC-NLPIR/Arbor

A generalist autonomous research agent — runs experiments, researches, and iteratively optimizes, autonomously.

What it solves

Arbor is an autonomous research agent designed for general-purpose optimization. It solves the problem of inefficient, "forgetful" AI-driven experimentation by replacing linear trial-and-error with a structured hypothesis-tree framework. This allows the agent to optimize any task with a measurable metric—such as model training, data synthesis, or harness engineering—while ensuring that gains are validated on held-out data to prevent overfitting.

How it works

Arbor employs two cooperating agents: a Coordinator (the research director) and an Executor (the research engineer).

They operate in a six-step "arbor cycle":

  1. Observe: Analyze current results and failure modes.
  2. Ideate: Propose new hypotheses based on the Idea Tree.
  3. Select: Prioritize the most promising ideas.
  4. Dispatch: The Executor implements changes in an isolated git worktree and evaluates them on a dev split.
  5. Backpropagate: Record results and abstract insights upward through the tree so future ideas inherit them.
  6. Decide: Use held-out validation to decide whether to merge the gain into the trunk, prune the branch, or continue.

Arbor can be used as a native CLI, integrated into other coding agents (like Claude Code or Codex) via MCP, or as a standalone skill suite.

Who it’s for

It is built for researchers and developers who need to optimize complex software or ML systems and want a disciplined, automated approach to experimentation that maintains a history of what worked and what failed.

Highlights

  • Hypothesis Tree: Maintains a structured history of research directions and concrete methods to avoid repeating mistakes.
  • Experiment Discipline: Uses isolated git worktrees and a strict dev/test split to ensure only genuine improvements are merged.
  • Literature Grounding: Integrates with the alphaXiv API to check the novelty of ideas before spending compute.
  • Flexible Integration: Supports multiple LLM backends (Anthropic, OpenAI, DeepSeek, etc.) and can run keylessly inside other AI coding harnesses.
  • Steerable: Offers a live dashboard and multiple interaction modes (auto, direction, review, collaborative) for human-in-the-loop guidance.

Related

  • Project
    harbor-framework/harbor

    Harbor is a Python framework for running and scaling benchmarks of AI agents and language models. Installable via `pip`/`uv`, it provides a CLI (`harbor run`) that lets you specify a dataset (e.g., Terminal‑Bench‑2.0), an agent (Claude‑Code, OpenHands, etc.), and a model, then executes the benchmark locally or on cloud providers (Daytona, Modal, etc.) with configurable parallelism. It also supports creating custom benchmarks and generating RL roll‑outs.

  • Project
    WILLOSCAR/research-units-pipeline-skills

    A framework for agent-assisted research that transforms research goals into verifiable, resumable runs with a full evidence trail and pinned pipeline contracts.

  • Project
    ARA-Labs/Agent-Native-Research-Artifact

    ARA‑Labs provides a set of seven “Agent Skills” that let AI research agents automatically record, verify, visualise and share every step of an experiment in a structured, cross‑layer artifact format, enabling reproducible, auditable autonomous research.

  • Project
    evo-hq/evo

    An autoresearch orchestrator that automates code optimization by running parallel AI-driven experiments in a loop to iteratively improve codebase metrics.

  • Project
    sapientinc/PRAXIST

    An autonomous research system that turns runnable, measurable projects into continuous, evidence-driven research runs to find optimal solutions through parallel agent exploration.